Autonomous Dual-Arm Robotics for Energized Electric Distribution Work

1. Engineering and Worker Safety Need

Electric utilities in the United States maintain energized distribution circuits by placing a trained human being inside arm's reach of a conductor at 4 to 35 kV. The work is skilled, it is governed by a mature standards regime, and it kills people at a rate almost no other occupation approaches.

Compiling Bureau of Labor Statistics Census of Fatal Occupational Injuries data for 2011 through 2024, the Electrical Safety Foundation International reports an electrical-cause fatality rate of 6.01 per 100,000 workers for electrical power-line installers and repairers, against 0.11 per 100,000 across all occupations. The scope qualifier is load-bearing and is stated here so that it travels with the number: both sides of that comparison are electrical-cause rates, which is what makes the comparison valid. It is not the whole hazard. Falls from elevation and vehicle incidents also kill lineworkers, so the trade's all-cause occupational fatality rate is materially higher than 6.01. That figure is not quoted here because the BLS tables carrying it were not retrievable to verification at the time of writing, and an unverified number does not belong in a research brief. The direction of the omission is worth stating: it understates the case, since a machine that removes the worker from the pole top addresses fall and struck-by exposure alongside electrical contact.

A second national figure is frequently placed beside these and should not be. Over the same period 2,070 of 70,276 recorded occupational fatalities were electrical, and contact with an overhead power line accounts for 49% of those electrical fatalities, ahead of unexpected contact with energy at 20% and nearby energized equipment at 12%. The denominator is electrical fatalities across all occupations, not lineworker deaths. A large share of that 49% is roofers, tree workers, crane and dump-truck operators, painters and sign installers making inadvertent contact. The figure establishes that the energized overhead conductor is the most lethal single object in American workplaces. It does not establish that half of lineworker deaths are overhead contact, and no such claim is made here.

Non-fatal exposure appears to be rising. ESFI records 5,180 non-fatal electrical injuries involving days away from work across 2023 and 2024, against 3,260 across 2021 and 2022. A 59% step in a mature BLS series is large enough that a case-classification or methodology change is at least as plausible an explanation as a real change in underlying risk, and the figure is reported here with that caveat attached rather than as an established trend. The workforce is meanwhile growing: BLS counts 127,400 electrical power-line installers and repairers employed in 2024 at a median wage of $92,560, with 7% growth projected through 2034, an additional 8,400 workers entering the same hazard.

The standard of practice offers three energized methods, all manual. Rubber glove work is performed from an insulated aerial bucket with the worker at ground potential and insulated from the line. Hot-stick work is performed at a distance using insulated tools. Barehand or equipotential work bonds the worker to line potential so that no current flows through the body. All three are governed by OSHA 29 CFR 1910.269, whose Table R-6 fixes minimum approach distances by voltage class, and by IEEE Standard 516, the guide for maintenance methods on energized power lines. Each is a legitimate engineering control and each is a highly trained craft. None of them removes the person from the hazard; they manage the person's presence within it.

The alternative, de-energizing the circuit, is being foreclosed by load growth. Planned outages are costly, poorly received by regulators and customers, and increasingly difficult to schedule on a system running closer to its limits. The result is that energized work is expanding, not contracting.

A second population is affected by the same asset condition. Reviewing 98 years of California fire-cause records on state-protected lands and 107 years on federal lands, Keeley and Syphard found that although ignition sources have declined markedly in recent decades, powerline ignitions are the notable exception to that decline, and identified reducing this source as "one important avenue for future fire-hazard reduction" (International Journal of Wildland Fire, 2018). Conductor and hardware condition drives that ignition risk, and conductor and hardware condition is precisely what energized maintenance exists to correct. A capability that makes energized maintenance safer and more frequent acts on worker mortality and on wildfire ignition through the same mechanism.

The asset base fixes the scale of the exposure. The United States operates approximately 600,000 miles of transmission line, of which about 240,000 miles are high voltage, and more than 5.5 million miles of local distribution line carried on over 180 million poles (ASCE, 2021 Infrastructure Report Card: Energy, drawing on Department of Energy data).

2. State of the Art

Four research and engineering communities work adjacent to this problem. They have not converged, and the shape of the non-convergence determines what a new programme should attempt.

Live-line robotic access, Hydro-Québec Research Institute (IREQ), Varennes, Québec. Serge Montambault and Nicolas Pouliot led the development of LineScout from 2006, a tracked platform that installs on an energized transmission conductor and travels along it, clearing obstacles including insulator strings, spacers and vibration dampers by a compliant re-gripping manoeuvre rather than by detaching. Their methodology is unusual for a robotics group and is the reason the work matters: they iterated against operational deployments rather than laboratory trials, publishing lessons on transportation and installation logistics, onboard energy management and odometry correction drawn from field use. The successor platform LineRanger addresses bundled conductors. IREQ licensed the technology to utilities including National Grid in the United Kingdom. The ceiling of this line of work is deliberate and structural: these are sensing platforms. They carry cameras and instruments, not manipulators, and cannot alter the state of the asset they traverse.

Autonomous live-line manipulation, Chinese grid research ecosystem. The only groups to have put autonomous manipulators onto energized distribution conductors are affiliated with China's State Grid research institutes and partner universities. Jiabo Feng and Weijun Zhang built an autonomous live-line maintenance robot for 10 kV overhead lines and took it to field experiments. Their architecture is a dynamic hierarchical task planner that decomposes an operation into directly executable instruction sequences, coupled to a cable recognition and positioning module engineered specifically to survive outdoor sunlight and cluttered backgrounds, plus live-line motion planning and a virtual environment for operator telepresence. Zhiqiang Wang's DARLOS took a different attack: a ground-up mechanical redesign to bring a 12-degree-of-freedom dual-arm system under the weight threshold at which it can ride existing aerial work platforms, with a quick-swap tool changer for multi-task capability. Zichen Tian and colleagues addressed a specific repair, broken strands on overhead ground wires, with two cooperating robots using an outer-fixation and inner-rotation architecture. Shuai Pei and colleagues took the electromagnetic environment itself as the object of study, modelling the field around a robot entering equipotential state on a 500 kV line, building an equivalent circuit for the transition and validating it experimentally.

Robotic contact work on de-energized utility assets, United States. ULC Technologies (Hauppauge, New York), working with Con Edison, developed a Cable Splicing Machine that clamps onto an uncut medium-voltage cable in a field structure and executes autonomous preparation operations using electrical actuators for coordinated motion. Field trials began in 2024 and the system received a 2024 RBR50 Robotics Innovation Award. The same partnership is developing a semi-autonomous mobile robot for racking medium-voltage circuit breakers in indoor distribution substations. This is genuine autonomous manipulation on utility hardware under a utility's own acceptance regime, and it is performed on de-energized assets.

Aerial live-line access, United States. LineBird, Inc. (Ashland, Virginia), founded by Michael Beiro, is developing unmanned aerial payload systems that deliver tooling to energized lines without a manned helicopter or bucket truck crew. The National Science Foundation has funded the approach through SBIR Phase I and Phase II.

Where the gap sits. The United States has robotic contact work on dead assets and aerial access to live ones. It does not have autonomous contact manipulation on energized conductors. That capability has been demonstrated, in the field, at distribution voltage, by groups working against State Grid construction standards, Chinese voltage classes and a vertically integrated utility structure. None of it has been ported to United States distribution hardware, none of it has been qualified against OSHA 1910.269 or IEEE 516, and the control architectures in every published system are hierarchical symbolic planners whose authors document the need for operator correction and hand-engineered perception. Learning-based manipulation, which is the dominant paradigm for contact-rich robotic tasks elsewhere, has not been applied to this domain at all.

3. Foundational Research

4. Competitive Landscape

Any single count of players in this space is an artifact of where the definition is drawn, so no single count is offered. The population is reported at three nested radii and the number rises as the definition loosens. Radius 1, any robot making sustained contact with an energized overhead conductor in North America: at least four. Hydro-Québec IREQ's LineScout and LineRanger traverse energized transmission conductors and are licensed to utilities including National Grid in the United Kingdom. FulcrumAir with Preformed Line Products installs BIRD-FLIGHT diverters and conductor spacers directly onto energized conductors, reporting the first robotic installation of PLP spacers and diverters in September 2023 and roughly 15,000 diverters along a 75-mile 345 kV line at the High Banks Wind Project in Kansas, with the system rated for installation on 240 kV energized conductors. Bombyx, developed for Meta by ULC Robotics and licensed in 2023 to HiBot, wraps fiber-optic cable onto medium-voltage distribution conductors while they remain energized, autonomously crossing insulators, taps and support structures at roughly 30 lb while carrying about 1 km of fiber. LineBird delivers tooling payloads for live-line access by unmanned aerial system. Radius 2, contact plus a change to the state of the line: three, as inspection-only platforms drop out. That is at the boundary of the conventional not-yet-commoditized threshold rather than comfortably beneath it, and this brief reports that rather than tightening the definition one further notch to obtain a better answer. Radius 3, dexterous multi-task autonomous manipulation on energized United States distribution: no commercial entrant identified. Every radius-2 system is a single-purpose payload placer that installs one class of hardware or strings one product; none carries a general manipulator assignable to a task it was not purpose-built for. Sarcos, now Palladyne AI, comes closest on hardware with the Guardian XT dexterous manipulator designed to mount on bucket trucks and boom lifts with power utilities among its named target industries, but it is teleoperated rather than autonomous and is not a live-line-qualified product.

What the three radii jointly establish matters more to a funder than the count itself. The claim that this space is untouched would be false and is not made. The accurate and more useful claim is that single-purpose robotic hardware placement on energized lines is a real, growing commercial activity with several credible entrants, while general-purpose autonomous manipulation on energized United States distribution has none. The first fact is the strongest available evidence for the feasibility of the second, because it demonstrates that utilities will accept robots working in contact with energized conductors once a safety case is made for a bounded task. The research problem is what happens when the task is no longer bounded.

The reason commercial solutions do not address the research problem is that a single-purpose placer can make a far narrower safety argument than a general manipulator. A machine that only ever installs one diverter on one conductor geometry has an enumerable failure space; a machine that will be handed arbitrary maintenance tasks does not, and no acceptance framework exists for the second case. That is a research and standards problem before it is a product problem, which is precisely why it has not been solved commercially. It is also why the commercially crowded layer, inspection, is not where a research programme should enter: Future Market Insights sizes powerline inspection robots at $760.6 million in 2026 growing to $3,185.8 million by 2036 at 15.4% CAGR, a healthy market with no research question left in it.

5. Addressable Scope and Public Benefit

A funder assessing this direction should see both the downstream economic scale and the public benefit that justifies the grant, and both are computed here from sourced inputs with every assumption labelled.

Scale of the domain. Investor-owned electric companies were projected to invest $208 billion in 2025, the highest figure recorded, comprising $66.5 billion in distribution and $39.9 billion in transmission, with more than $1.1 trillion projected across 2025 through 2029 (Edison Electric Institute, Industry Capital Expenditures, September 2025). This is the spending envelope the capability sits inside; it is not the addressable market and is not presented as one.

Bottom-up addressable value. Each factor is marked SOURCED or ASSUMPTION, and unsourced factors are bracketed rather than point-estimated because no published figure exists for the energized share of distribution field labour.

Employed electrical power-line installers and repairers: 127,400 (SOURCED: BLS Occupational Outlook Handbook, 2024). Median annual wage: $92,560 (SOURCED: BLS, 2024), giving a direct wage base of $11.79 billion. Fully loaded multiplier for benefits, vehicle, tooling and overhead: 1.4x to 1.6x (ASSUMPTION, bracketed; typical utility employee burden. A 2.0x figure would be a contractor billable rate rather than an employer cost and is not used), giving a field labour base of approximately $16.5 billion to $18.9 billion per year; median wage understates the mean, so this base is conservative. Share of that labour performing energized rather than de-energized construction, new service and storm restoration: 10% to 20% (ASSUMPTION, bracketed). Energized-work labour: $1.7 billion to $3.8 billion per year. Share of energized task time a dual-arm system can absorb at maturity, with access, set-up, switching and site control remaining human: 20% to 40% (ASSUMPTION, bracketed). Serviceable task value: $330 million to $1.5 billion per year. One caution belongs inside the calculation rather than after it: a material share of lineworker hours is capitalised against construction projects rather than expensed as maintenance, so the labour base is not a single homogeneous pool and the energized-maintenance subset is the part this calculation concerns.

Top-down reference points, and an honest statement of what they are not. Two figures circulate around this space and neither is an independent validation of the estimate above. The National Science Foundation's Phase II abstract for LineBird states the technology targets “an addressable market valued at over $16 billion in overhead line maintenance and smart grid solutions” (NSF Award #2450659). That figure originates in a company's own SBIR application, is unaudited, and spans a far broader scope; it is a vendor-asserted number reproduced in a government abstract, not a market study. Future Market Insights sizes the powerline inspection robot market at $760.6 million in 2026 rising to $3,185.8 million by 2036 at 15.4% CAGR, which measures an adjacent layer; using an inspection figure to validate a contact-work market would be a category error. Jointly they establish something narrower but real: utilities already procure robotic systems at the hundreds-of-millions scale, and a credible operator in the niche believes the surrounding opportunity is measured in billions. The bottom-up range is the estimate; these are context.

Public benefit at scale, which is the return on a public research dollar. Contact with an overhead power line causes 49% of workplace electrical fatalities. A capability that removes the human from inside the minimum approach distance acts directly on the largest single category of electrical death in American workplaces, in an occupation of 127,400 people that is growing 7% per decade and whose non-fatal injury count rose 59% between the last two reporting periods. The second benefit runs through asset condition: powerline ignitions are the one wildfire ignition source that has not declined (Keeley and Syphard, 2018), and more frequent, cheaper energized maintenance is a direct lever on conductor and hardware condition. Both benefits accrue across 5.5 million miles of distribution line and 180 million poles, and neither requires the technology to reach the full serviceable market to be realised.

Cost recovery, this sector's analogue of clinical reimbursement. CPT and HCPCS codes govern reimbursement for a clinical device and have no application here; the substitution is stated rather than silently omitted. The governing mechanism is regulated cost recovery. Capital equipment enters the utility rate base and earns an authorised rate of return set by the state public utility commission, while operations and maintenance is expensed and recovered without a return. FERC's transmission rate incentives are deliberately not cited, because FERC has no rate jurisdiction over retail distribution, which is where this opportunity sits. The relevant state-level example is California, where wildfire mitigation plan spending is recoverable under California Public Utilities Code section 8386, with plan review sitting since 2021 with the Office of Energy Infrastructure Safety rather than the CPUC, under the framework established by AB 1054. Section 6 explains why the capital-versus-expense distinction, though real, is not the reason incumbents have not moved.

6. Research Gap and HHA Contribution

What the cited researchers have not done. Four things, stated precisely.

First, no published system uses a learned control policy. Feng and Zhang's architecture is a dynamic hierarchical task planner, which is to say an enumeration of the situations its designers anticipated, and the paper documents both consequences: residual operator correction, and perception that had to be engineered specifically against sunlight and background clutter. A symbolic planner is the correct engineering choice for a well-specified environment. Pole-top distribution is not one. Conductor sag and sway change the approach geometry between one attempt and the next, hardware varies between utilities and across decades of construction vintage, and weather changes the scene continuously.

Second, no published system treats the electromagnetic environment as a control input. Pei and colleagues measured a 100 to 200 A arc with energy below 100 MHz occurring at the equipotential transition and recommended routing the radio around it. Routing around a disturbance is not the same as sensing it. A controller that observes field strength and adapts its approach trajectory has not been built.

Third, none of this work has been specified against United States distribution construction. The demonstrated systems address 10 kV Chinese distribution and 500 kV transmission. United States primary distribution runs 4 to 35 kV on crossarm and armless constructions with hardware, cutouts, connectors and clearances that differ materially, and the entire body of work is silent on OSHA 1910.269 Table R-6 and IEEE 516.

Fourth, there is no acceptance framework. No United States protocol exists by which a utility or a state commission can accept a robot operating inside the minimum approach distance. There is no test method, no instrumentation standard, and no evidentiary threshold.

The precise technical gap between published results and a deployable system is therefore the substitution of a learned, safety-filtered manipulation policy for a symbolic planner, validated on United States distribution hardware, with the electromagnetic environment admitted as state, wrapped in an acceptance case a utility safety organisation will sign.

What HHA's research programme would do differently, as methodology rather than as goal. The control core is a policy trained by off-policy reinforcement learning, specifically Soft Actor-Critic, selected over on-policy alternatives for sample efficiency on contact-rich tasks and over pure behaviour cloning because the demonstration set for energized work is necessarily small. The policy is warm-started by behaviour cloning from teleoperated demonstrations captured on a de-energized full-scale mock-up, then improved in simulation. The state space comprises global-shutter stereo RGB-D at 30 Hz, an event camera whose microsecond temporal resolution and high dynamic range address the exact sunlight failure Feng and Zhang document, six-axis force-torque sensing at each wrist at 1 kHz, joint encoders at 1 kHz, and an electric-field probe sampled at 10 kHz supplying the electromagnetic observable that no prior system carries. The action space is bounded joint velocity across 12 degrees of freedom, with end-effector speed limited to 0.05 m/s inside one metre of an energized conductor. The reward is task completion, penalised for contact force above a hardware-derived threshold and for elapsed time, with approach-distance violation handled outside the reward entirely.

That last point is the safety architecture and it is deliberate. The OSHA Table R-6 minimum approach distance is enforced by a control barrier function acting as a filter on the policy's output, not by a penalty term in the reward. A reward penalty makes a violation expensive; a barrier function makes it kinematically unreachable. A learned policy that has merely been taught that approach violations are costly is not a system a utility safety organisation can accept, and the distinction is the difference between a demonstration and a qualifiable machine. Training uses domain randomisation over illumination, conductor sag, wind-induced sway, hardware vintage and surface condition, with the electrostatic field solved offline by finite element analysis and injected into the simulation. Validation proceeds from simulation to a de-energized full-scale pole yard of the kind utilities operate as hot-line schools, and only then to a utility test facility under energization.

Why HHA is positioned for this. The gap decomposes into three capabilities that rarely co-occur, and each maps to a named team member in Section 9. Learned control under hard safety constraints and the evaluation methodology to characterise tail behaviour rather than average behaviour is Haedar Hadi's. The applied physics of the high-voltage environment, the instrumentation to observe it, and the experimental design that renders a safety claim falsifiable to a utility's own safety organisation is Hass Dhia's. Design for manufacturability, without which the output is a prototype rather than a capability, is Ahmed's, and Section 9 argues that this third capability is the one whose absence most often strands work of exactly this type.

Why the originating labs and the incumbents have not closed the gap, which is the question a reviewer should ask. The answer is not scientific, and the version of it that circulates is wrong, so it is worth stating carefully.

The naive account holds that rate regulation punishes labour-saving tooling, because a utility earns an authorised return on rate-based capital while merely recovering operations and maintenance expense. That reasoning does not survive contact with the mechanism it invokes. A robot is capital. A utility that buys a fleet of live-line robots capitalises them and earns its authorised return across their depreciable life, while the crew hours displaced sit substantially in expense. Under cost-of-service regulation that substitution is attractive rather than unattractive, and the classical Averch-Johnson result predicts a bias toward exactly it. Anyone who has read a rate case notices the error at once, so the correct account is different.

The real constraint is retention, not classification. Under cost-of-service, an efficiency saving passes through to ratepayers at the next rate case, so the utility retains it only across regulatory lag, typically one to three years. The payback horizon for a productivity investment is truncated no matter how the asset is booked. The consequence is not that utilities cannot buy robots; it is that the case cannot be denominated in wages. It has to be denominated in what regulators reward: SAIDI and SAIFI reliability performance, storm restoration and mutual-aid surge capacity, and, since 2017, wildfire ignition risk, which has become the dominant capital allocator in the western United States.

A second barrier is larger and is usually missed. Most energized line work in the United States is not performed by utility crews but bought from contractors, principally Quanta Services, MYR Group, MasTec and Pike, whose revenue model is billable crew hours. A tool whose value proposition is removing crew hours is directly adverse to the profit and loss of the party who would operate it. That is a sharper adoption barrier than anything in the regulatory structure, and it determines who the buyer is: the asset owner's reliability and wildfire-mitigation programme, not the contractor's operations budget.

Two further reasons compound. Collective bargaining, principally with the IBEW, makes automation framed as headcount reduction politically costly, a second independent reason hazard removal is the only credible framing. And the institutions with the deepest capability are structurally disinclined: Hydro-Québec's IREQ is a utility research arm without a commercialization mandate, and the Chinese grid institutes have no incentive to requalify their work for a foreign regulatory regime whose rules differ from the ones their systems were built against.

One inference this brief declines to draw. It is tempting to read the DOE Grid Resilience and Innovation Partnerships portfolio, roughly $4.2 billion across 46 projects announced 18 October 2024 within a $10.5 billion programme, as evidence that funders undervalue the robotic labour layer. That reading is unsound. GRIP is a deployment programme whose announcements specify resilience and smart grid topic areas and whose awards carry cost-share and shovel-readiness requirements that no TRL 4 research effort could meet. The absence of robotics there is a programme-design selection effect, not a revealed preference, and Section 7 treats it accordingly.

A university laboratory that closed the science would still face the qualification problem, which is neither publishable nor fundable as science. That is precisely the kind of work an applied research institute with a manufacturing capability exists to do, and it is why this gap has stayed open while the surrounding field advanced.

7. Comparable Funded Projects

LineBird, Inc. (Ashland, Virginia). Principal Investigator Michael Beiro. NSF SBIR Phase I, Award #2136680, $256,000, 15 September 2022 to 31 December 2024. The award funds unmanned aerial systems performing inspections and repairs on live transmission equipment without manned helicopter or bucket truck crews. NSF's abstract states the aim of “reducing the time linemen are in harm's way” and anticipates more frequent routine inspection and improved grid resilience. This is the closest direct precedent for the premise of the present direction, and its existence establishes that a federal agency accepts removing humans from live-line work as a fundable objective. Its approach, aerial payload delivery, is complementary rather than competing with a manipulator riding an insulated aerial platform.

LineBird, Inc. NSF SBIR Phase II, Award #2450659, $312,496, 1 September 2026 to 31 August 2028. Phase II funds commercialization of the same live-line access system. The abstract cites “an addressable market valued at over $16 billion in overhead line maintenance and smart grid solutions” and names workplace safety, operational cost reduction and grid resilience as the broader impacts. That a Phase I converted to Phase II is a stronger signal than either award alone: NSF reviewed the Phase I results and judged the commercialization case sound. It also supplies an independent, government-published market figure for Section 5.

DOE Grid Resilience and Innovation Partnerships (GRIP), approximately $4.2 billion across 46 projects in 47 states and the District of Columbia, announced 18 October 2024, within a $10.5 billion programme; a second funding opportunity of up to $3.9 billion has been announced. GRIP is the largest current federal vehicle for grid resilience and its scope includes substation automation and digitization. Its relevance is that it establishes the problem domain commands funding at a scale dwarfing anything this direction would request. Its award composition does not establish that funders undervalue the robotic labour layer: GRIP is a deployment programme whose cost-share and shovel-readiness requirements structurally exclude TRL 4 research, so the absence of robotics in the portfolio is a selection effect of programme design rather than a revealed preference, and reading it otherwise would be reverse-engineering evidence to fit a thesis.

Con Edison with ULC Technologies, Cable Splicing Machine, utility-funded, field trials from 2024, 2024 RBR50 Robotics Innovation Award; amount not publicly disclosed. Not a federal grant, and included because it answers a different question than the grants do: whether a major investor-owned utility will fund, host and field-trial autonomous robotic manipulation on its own distribution assets. It will, and it has, on de-energized cable. This materially de-risks the utility-partnership assumption on which any energized programme depends.

What the pattern establishes, and the one gap in it. Funders are spending at very large scale on grid resilience and NSF has twice backed a company whose thesis is removing humans from live-line work, converting a Phase I into a Phase II, so the premise is demonstrably fundable. What is absent is any substantial federal research award treating autonomous contact manipulation on energized conductors as a research problem. A search of the NSF award database returns exactly two awards for the term “live-line,” both to LineBird, while a search for “substation” returns a portfolio dominated by cybersecurity and cyber-physical systems with no robotics component. That absence is a claim about the portfolio and not about the instrument: the same search method returned the LineBird awards and genuine substation awards correctly, which is the positive control that makes the negative result meaningful. Read against the standard diagnostic, this is not a field that is too early, since field results on energized lines exist, and it is not a problem funders find uncompelling. It is an opportunity between programme boundaries, where robotics research programmes do not typically fund grid applications and grid deployment programmes cannot fund pre-deployment research, and a proposal must be framed deliberately for one or the other rather than assuming a natural home.

8. Opportunity Assessment

TRL evidence chain

The integrated capability is TRL 4, on a component chain spanning 3 to 7. TRL attaches to a specific system in a specific application and environment, so the high component numbers below do not transfer to the integrated configuration.

Conductor mobility and deployment, TRL 6 to 7: more than 20 LineScout field deployments on energized transmission lines (Pouliot and Montambault, 2012) plus continuing commercial use. Single-purpose energized hardware placement, TRL 6 to 7: FulcrumAir and PLP, roughly 15,000 diverters on a 75-mile 345 kV line in 2023, and Bombyx on energized medium-voltage distribution. Dual-arm manipulator hardware, TRL 5: DARLOS, emulated environment only (Wang, 2024). Assisted manipulation on energized distribution, TRL 5: field experiments at 10 kV (Feng and Zhang, 2021). Perception and manipulation under arc-discharge EMI, TRL 3 to 4: Pei et al. (2025) characterises the disturbance; no published system controls against it. Dielectric qualification to United States standards, TRL 3: not addressed anywhere in the cited literature. Integrated autonomous execution on United States distribution, TRL 3 to 4: Tian et al. (2026) is explicitly laboratory. The lowest component governs, so the integrated system is TRL 4, gated by dielectric qualification and by autonomy under arc-discharge EMI.

One qualifier that must travel with the Feng and Zhang result. Their own phrasing is that the robot completes most operations automatically “requiring only a small amount of correction and operation.” That describes assisted operation with a human in the loop, not autonomy, and this brief does not cite it as field-proven autonomy. It is field-proven useful energized robotic work at competitive speed, a different and still significant claim. The 10 kV figure is also a Chinese distribution class; United States primary distribution runs 4.16, 12.47, 13.2, 24.9 and 34.5 kV with different phase spacing, crossarm geometry, conductor types and construction standards, so the result is real but the transfer is not free.

Technical risks as research questions

Can a learned policy hold conductor pose estimation under full-sun outdoor illumination where a hand-engineered pipeline required special treatment? Feng and Zhang solved this by engineering; the question is whether event-based sensing plus domain-randomised training solves it by learning, which would make the result portable across hardware and utilities rather than tuned to one site. Go/no-go at month 9: conductor pose estimation better than 10 mm root-mean-square across a full solar cycle on an outdoor de-energized mock-up. If the threshold is missed, the fallback is a hybrid architecture retaining a geometric estimator with the learned policy operating on its residual.

Does admitting measured electric field strength into the policy state improve behaviour through the equipotential transition, or is shielding sufficient? This is the question Pei and colleagues' measurement raises and does not answer. The experiment is a two-arm comparison, policy with and without the field observable, over instrumented transitions. Go/no-go at month 15: zero control-loop faults across 100 instrumented equipotential transitions, with the field-observable arm showing a measurable reduction in trajectory deviation. A negative result is publishable and materially useful to the field, since it would establish shielding as sufficient.

Will a utility safety organisation accept a barrier-function safety argument? This is a research question about evidence rather than about engineering, and it is the one that cannot be retired in a laboratory. It is addressed by co-designing the acceptance protocol with a partner utility's safety organisation before hardware freeze, so the resulting protocol is the partner's rather than a vendor submission. The programme plan should state openly that this risk is not fully retireable by engineering.

Regulatory pathway

This is not a medical device and there is no FDA nexus, so the biomedical pathway questions are answered by substitution rather than omission. The governing instruments are OSHA 29 CFR 1910.269 for maintenance work, whose Table R-6 fixes minimum approach distances by voltage class, together with 29 CFR 1926 Subpart V, which governs construction work on distribution lines; the two were harmonised in 2014 but the applicability split still determines which rules bind a given task. The equipment standards a live-line engineer expects to see named are IEEE Standard 516 for maintenance methods on energized power lines, ASTM F711 for the fiberglass rod and tube used in live-line tools and booms, and ANSI/SAIA A92.2 for vehicle-mounted insulating aerial devices, which is the platform this system rides. State general orders apply on top: in California, GO 95 for overhead construction and GO 165 for inspection cycles. Any aerial component falls under FAA Part 107, with beyond-visual-line-of-sight operation requiring a waiver. Human-subjects IRB review is not triggered, since operators are employees performing their occupation rather than research subjects, though a partner utility's own internal review will apply.

The approach-distance question is subtler than it first appears, and it is the one a utility safety director asks in the second minute. Table R-6 minimum approach distances are written for employees. A robot is not an employee, so the standard does not straightforwardly bind the machine. What it binds is the human operator, the boom, the tether, the operator's position and any conductive path back to them, and the prior question is whether the work is classified as insulate-and-isolate or as equipotential. Treating the robot's own clearance as the whole regulatory question is the error that marks a team which has not sat with a safety organisation. The position taken here is that the approach distance should still be enforced kinematically on the manipulator as a conservative design constraint, while the qualification argument is built where the standard actually bites. Resolving that classification with the partner safety organisation is a month-one activity, not a month-eighteen one.

Locked versus adaptive control and the precedent that governs it. The distinction FDA draws for clinical algorithms applies here with equal force and no regulator in this sector has yet drawn it. A policy frozen at release can be qualified once against a fixed protocol. A policy that continues to learn on deployed hardware is no longer the machine that was qualified. This programme adopts a frozen policy at release governed by a documented change-control plan specifying what may be retrained, on what data, and what requalification each change triggers, with full decision-trajectory logging so any field event can be replayed against the released policy. The on-point governance instruments are the utility's own management-of-change process, NERC CIP where anything touches control systems, and the state general order revision process, and those are what an acceptance package must actually satisfy. FDA's Predetermined Change Control Plan for machine-learning-enabled device software, finalised December 2024, is cited here as a structural analogy only, because it is the most mature worked example of a regulator authorising bounded post-deployment algorithm change. It carries no jurisdiction here and is not offered as regulatory cover. Precedent systems already accepted into utility service, LineScout for energized transmission inspection and the Con Edison and ULC cable splicing machine for de-energized distribution work, demonstrate that utilities do accept robots onto their assets once a case is made; neither case has been made for energized contact manipulation.

Regulatory position as a moat rather than a barrier. Because no acceptance framework exists, the first system qualified for energized contact work defines the test protocol, the instrumentation and the evidentiary standard every follower must then satisfy. The two to four years required to establish it is a barrier to entry that scales with neither capital nor headcount, which is what makes an early, deliberately public contribution to that framework strategically valuable and consistent with HHA's open-methodology stance.

Proposed experimental approach, first six months

Months 1 to 3: build the simulation environment with finite-element electrostatic fields injected, and instrument a de-energized full-scale distribution mock-up at a partner hot-line school. Capture the teleoperated demonstration set for behaviour-cloning warm start. In parallel, Ahmed's team establishes the manufacturing baseline against Wang's sub-150 kg envelope, and the partner utility's safety organisation is engaged to begin drafting the acceptance protocol. Months 4 to 6: train the first policies under domain randomisation, run the barrier-function safety filter against adversarial trajectory injection in simulation, and execute the first de-energized mock-up trials on a single task, most likely a connector or cutout operation chosen with the partner. The month 6 gate is a policy completing that single task autonomously on the de-energized mock-up at a success rate above 90% across 100 trials with zero barrier-function violations.

9. Team Fit

Hass Dhia — Co-Principal Investigator. MS Biomedical Sciences (Wayne State University School of Medicine), medical school background including service as an anatomy teaching assistant, and an AI infrastructure architect with a physical-sciences foundation spanning physics, chemistry, thermodynamics and fluid dynamics. Mapping to components: the applied physics of the high-voltage working environment, including electric field modelling and characterisation of the 100 to 200 A arc that Pei and colleagues measured at the equipotential transition; specification and integration of the electric-field probe and the sensor-fusion stack that admits that measurement into the control policy's state; and experimental design for the two-arm field-observable comparison and the de-energized mock-up trials, structured so that a safety claim is falsifiable rather than merely asserted. The methodological through-line from his training is quantitative measurement of a physical system under uncertainty and the design of experiments that can disconfirm a safety hypothesis, which is the same discipline a utility safety organisation applies, expressed in a different domain. He carries the OSHA 1910.269 and IEEE 516 framing of the acceptance case.

Haedar Hadi — Lead Principal Investigator. MS Computer Science, Boston University, Information Systems focus, with cloud and database architecture experience. Mapping to components: the learned control core, comprising the Soft Actor-Critic policy, the behaviour-cloning warm start from teleoperated demonstrations, and the domain-randomisation regime over illumination, conductor sag, sway and hardware vintage; the control-barrier-function safety filter that renders minimum-approach-distance violation kinematically unreachable rather than merely costly; and the evaluation methodology and benchmark design, which in this domain must characterise tail behaviour rather than average performance, since a policy that succeeds 99% of the time and fails catastrophically in the remaining 1% is not acceptable at any average. The scalable compute and data infrastructure for the training regime and for full decision-trajectory logging is his as well, and that logging is what makes the frozen-policy change-control package auditable.

Ahmed — Director of Manufacturing. Mapping to components: design for manufacturability of the dual-arm structure against the sub-150 kg envelope that Wang identifies as the threshold for compatibility with the aerial work platforms utilities already own; tolerance analysis across 12 degrees of freedom, where accumulated error at the end effector is what determines whether a barrier-function margin is achievable in hardware rather than only in simulation; qualification of dielectric materials at distribution voltages; ingress-protected actuator selection for weather and wash-down exposure; field-serviceable modular architecture; and the quality system and inspection plan a utility will require before accepting equipment onto its assets.

The lab-to-production bridge. Most research proposals end at “it works in the lab.” This proposal includes explicit DFM milestones at every phase, ensuring that prototype decisions consider production scaling, tolerance analysis, and quality systems from day one. This addresses the valley of death between TRL 4-5 prototypes and TRL 7+ deployable systems, the gap where most funded research stalls. In this specific domain the argument is not generic. Every system cited in Section 3 is a hand-built prototype, and the one result that most determines commercial adoptability, Wang's sub-150 kg system weight, is a manufacturing achievement rather than a robotics one. A programme that defers manufacturing until the science is settled will discover at TRL 5 that its tolerance stack cannot hold the safety margin its policy assumes, which is a failure that cannot be corrected in software.

What this team does not have, and how funding addresses it. HHA holds neither a utility partnership nor an energized test facility, and neither can be substituted for. The programme is therefore structured around a partner utility recruited before hardware freeze, with the acceptance protocol co-authored by that partner's safety organisation, and budget is allocated to subcontract energized testing to a utility or independent high-voltage laboratory rather than to build such a facility. The team also lacks a journeyman lineworker's craft knowledge, which is not an academic input and cannot be read out of the literature; the programme budgets for a subject-matter expert on subcontract from month one, both to inform task selection and to keep the demonstration set honest. HHA is positioned as the computational and manufacturing enabler for a capability that utilities and high-voltage laboratories will validate, not as a competitor to either.

10. Recommended Next Steps

Target funder programmes

NSF Small Business Innovation Research / Small Business Technology Transfer, Advanced Manufacturing and Robotics topic. The strongest first target on precedent: NSF has funded live-line access twice under this mechanism, and Phase I converted to Phase II, so the reviewing panel has already accepted the premise. An STTR structure suits a team combining research and manufacturing capability. Comparable Phase I awards in this line ran $256,000 and Phase II $312,496.

NSF Foundational Research in Robotics (FRR), and the Cyber-Physical Systems programme. The learned-policy-plus-barrier-function safety architecture under a task-correlated electromagnetic disturbance is a foundational robotics contribution independent of its application, and CPS is the natural home for the safety-filter formalism. This is the route for the science if the SBIR route carries the translation.

DOE Office of Electricity, and the Grid Resilience and Innovation Partnerships second funding opportunity (up to $3.9 billion announced). The framing here must be resilience and wildfire mitigation rather than robotics, since GRIP funds grid outcomes. A utility partner is effectively a precondition for this route and it is the largest available vehicle.

Electric Power Research Institute collaborative programmes. Not a grant funder in the federal sense, but the mechanism by which a technology enters utility acceptance and the fastest route to the partner utility every other application depends on. This should be pursued first in time even though it carries the least money.

Estimated funding range. On the comparable awards in Section 7, a Phase I equivalent sits at $250,000 to $400,000 for twelve months, and a Phase II or an NSF research award at $1.0 million to $2.0 million over twenty-four months. The twenty-four-month programme below is scoped against the upper figure and is the minimum viable preclinical-equivalent programme: below roughly $1.2 million the energized-facility subcontract and the manufacturing track cannot both be funded, and dropping either produces a result that cannot progress.

Twenty-four month milestone plan

• M1–3 R&D: Simulation environment built with finite-element electrostatic fields injected. De-energized full-scale distribution mock-up instrumented at a partner hot-line school. Teleoperated demonstration set captured for behaviour-cloning warm start.
• M1–3 Manufacturing (Ahmed), starting month 1: Manufacturing baseline established against the sub-150 kg envelope. Tolerance budget allocated across the 12-DoF structure and reconciled against the barrier-function margin the control design assumes. Dielectric material candidates identified for distribution-voltage qualification.
• M1–4 Regulatory and partnership (parallel track): Partner utility recruited and its safety organisation engaged. Acceptance protocol drafting begun, co-authored rather than vendor-submitted. Journeyman lineworker subject-matter expert onboarded on subcontract.
• M4–6 R&D: First policies trained under domain randomisation. Barrier-function filter tested against adversarial trajectory injection in simulation. Month 6 gate: single task completed autonomously on the de-energized mock-up at above 90% success across 100 trials with zero barrier-function violations.
• M7–9 R&D: Event-camera perception integrated and evaluated across a full solar cycle. Month 9 gate: conductor pose estimation better than 10 mm RMS under full-sun outdoor conditions, or fall back to the hybrid geometric-plus-residual architecture and document the decision.
• M9–15 R&D: Electric-field probe integrated. Two-arm comparison run, policy with and without the field observable, across instrumented equipotential transitions. Month 15 gate: zero control-loop faults across 100 transitions. A null result on the field-observable arm is recorded and published rather than buried.
• M10–18 Manufacturing (Ahmed): Second-iteration arm structure built to production tolerances. Ingress-protection and wash-down qualification. Field-serviceable modular breakdown validated. Inspection plan and quality documentation drafted to a standard a utility will accept.
• M15–22 Regulatory (parallel track): Acceptance protocol finalised with the partner safety organisation. Frozen-policy release package assembled with full decision-trajectory logging and a documented change-control plan structured on the PCCP model. Submitted to the partner utility for internal review.
• M18–24 R&D and programme: Energized trial at a utility or independent high-voltage test facility on the selected task, executed under the co-authored acceptance protocol. TRL 5 declared on that evidence or the gap documented and the follow-on scoped against it. Simulation environment, benchmark and acceptance protocol released openly per HHA's methodology stance. Phase 2 proposal prepared for multi-task capability and a second utility partner.