Technology
The Power Problem: How Humanoid Robots Manage Energy From Battery to Joint
Humanoid robots walk at 500 watts and carry 2.3 kWh batteries that last 2-4 hours. Until that ratio improves, the 8-hour factory shift remains out of reach. Here is how robot power architecture works and what will change it.
Every humanoid robot demo ends the same way. The robot walks, waves, lifts a box, sorts packages with uncanny dexterity — and then, somewhere between 90 minutes and four hours later, it stops. Not because of a software crash or a mechanical failure, but because the battery died. Power is the invisible bottleneck holding humanoid robotics back from commercial reality, and almost nobody talks about it.
The gap between what these machines can do and how long they can do it is stark. A human being walking at a brisk pace consumes roughly 100–200 watts of metabolic power. The brain — running the most sophisticated perception and motor-control system ever built — draws just 20 watts. A humanoid robot doing the same walk burns 400–600 watts, and its onboard AI chips add another 40–130 watts on top of that. The physics are brutal.
Understanding why humanoid robots have battery problems — and what engineers are doing about it — requires looking at the full power chain: from cell chemistry and pack architecture through power distribution, motor drives, and the compute tax that modern AI imposes on every cycle of motion.
AI-generated image
Key Stats
2.3 kWh
Optimus Gen 2 battery capacity
500W
Typical walking power draw
70–80%
Power consumed by motors
130W
NVIDIA Jetson Thor compute draw
How Much Power Does a Humanoid Robot Actually Use?
The numbers are more extreme than most people expect. Tesla Optimus Gen 2, one of the most closely watched platforms in the industry, draws approximately 100 watts at idle — just standing still, running sensors and control loops. Walking bumps that to roughly 500 watts. Reaching to lift a heavy object can spike demand to 1,000–2,000 watts as multiple actuators fire simultaneously at high torque.
The old Boston Dynamics Atlas — the hydraulic-powered research platform famous for backflips — offers an instructive extreme case. That robot carried a 3.7 kWh battery pack and lasted roughly one hour on a charge, implying an average draw of around 3,700 watts. Hydraulic actuation is spectacularly powerful but equally spectacularly inefficient. The transition to electric motors has been one of the most important engineering shifts in the field.
Breaking down where power goes in a modern electric humanoid reveals a clear hierarchy:
• Electric motors and actuators: 70–80% of total draw. Every joint — ankles, knees, hips, shoulders, elbows, wrists, fingers — has its own motor. Running 20–40 motors simultaneously during dynamic motion is the dominant load by a wide margin.
• Onboard compute: 15–20% of total draw. The NVIDIA Jetson Orin, a common choice for mid-tier platforms, draws 15–75 watts depending on load. The newer Jetson Thor — 40–130 watts, delivering 2,070 TOPS of AI compute — represents the premium end of current deployments.
• Sensors and communications: Under 50 watts combined. LiDAR, stereo cameras, IMUs, force-torque sensors, and Wi-Fi/5G radios all contribute, but their share of total draw is relatively modest.
The core tension: Humanoid robots need to be light enough to move efficiently, but the battery must be heavy enough to power hours of operation. These two requirements pull in exactly opposite directions — and every design team in the industry is wrestling with where to draw the line.
The Battery Pack: Chemistry, Weight, and the 1/8 Rule
The battery pack in a humanoid robot is not just a power source — it is a structural and balance constraint. Because the pack must sit in the torso to keep the robot center of mass stable during dynamic motion, its weight directly affects every joint in the body. Heavier packs increase the load on hip, knee, and ankle actuators during every step, which in turn increases power draw, which shrinks the runtime gains from adding more capacity. The industry has converged on a rough heuristic: battery mass should not exceed approximately 1/8 of total robot body weight.
Current platforms illustrate the tradeoffs clearly:
• Tesla Optimus Gen 2: 2.3 kWh capacity at approximately 52V nominal voltage. The robot weighs 57 kg. The pack sits near the torso center of mass to minimize dynamic imbalance during motion.
• Figure 01/02: 2.25 kWh pack with a claimed runtime of up to 5 hours under light-duty cycles. At approximately 70 kg total robot weight, Figure has more mass budget to work with and can afford a slightly heavier pack without the same penalty.
• Unitree H1: Dual 0.864 kWh batteries for a combined 1.73 kWh total. The H1 weighs under 50 kg and achieves 1–2 hours of active runtime. Its hot-swappable design compensates somewhat for the smaller pack.
At the cell level, current lithium-ion and lithium-polymer cells achieve energy densities of 230–300 Wh/kg. When assembled into a complete battery pack — with thermal management, BMS electronics, structural casing, wiring, and connectors — system-level density drops to 150–250 Wh/kg. That loss is the price of safety and longevity.
The main bus voltage has converged around 48V nominal across most commercial platforms. This level balances cable gauge requirements (higher voltage means lower current for the same power, enabling lighter wiring), motor drive efficiency, and safety margins for personnel working near the robots.
Platform Power Comparison
| Platform | Battery | Runtime | Avg Draw | Weight |
|---|---|---|---|---|
| Tesla Optimus Gen 2 | 2.3 kWh | ~2–4 hrs | ~500W avg | 57 kg |
| Figure 01/02 | 2.25 kWh | ~5 hrs claimed | ~450W avg | ~70 kg |
| Unitree H1 | 0.864 kWh | 1–2 hrs | ~500W avg | <50 kg |
| Boston Dynamics Atlas (hydraulic) | 3.7 kWh | ~1 hr | ~3,700W avg | 89 kg |
| Agility Digit | ~2 kWh | 2–4 hrs | ~500W avg | 65 kg |
Power Distribution: From Pack to Joint
Getting 500 watts from a torso-mounted battery pack to 30-plus motors distributed across limbs, hands, and a head is not a trivial engineering problem. Modern humanoid robots use a zonal power distribution architecture, with a central Power Distribution Unit (PDU) in the torso managing the main 48V bus and multiple DC-DC converters stepping voltage down for different subsystems.
A typical voltage hierarchy looks like this:
• 48V main bus: Leg and arm motor drives. These joints handle the highest forces and need the most headroom for regenerative braking capture during stepping.
• 24V bus: Hand and finger actuators. Dexterous manipulation requires precise low-power motors where 48V would over-specify the hardware and add unnecessary switching losses.
• 12V bus: Sensors, IMUs, force-torque sensors, cameras, LiDAR, audio systems. Low-power peripherals that need clean, stable voltage.
• Isolated compute rail: GPU modules, inference accelerators, and communications hardware run on dedicated isolated rails to prevent voltage noise from motor switching from corrupting compute signals.
Battery Management Systems from suppliers like Infineon handle cell balancing, state-of-charge estimation, temperature monitoring, and fault protection at the pack level. The motor drives themselves use Field-Oriented Control (FOC) — a technique that optimizes torque delivery per watt by precisely controlling the phase angle between rotor and stator magnetic fields. FOC implementation quality is a meaningful differentiator in motor efficiency across platforms.
At the transistor level, the industry is moving from silicon MOSFETs to Gallium Nitride (GaN) transistors. GaN switches faster, runs hotter, and has lower on-state resistance than silicon, translating directly to higher motor drive efficiency — often 2–5 percentage points better at the switching frequencies used in robot actuators. Over millions of switching cycles per hour, that margin compounds into meaningful runtime extension.
One often-overlooked challenge is the wiring harness itself. Cables running from torso to wrists and ankles must survive more than one million flexion cycles without failing. Routing, bend radius management, and connector selection are mechanical reliability problems that receive as much engineering attention as the electronics they serve.
Walking Efficiency: The Cost of Transport Metric
Roboticists measure walking efficiency using a dimensionless metric called the Cost of Transport (COT): the energy consumed divided by the product of body weight and distance traveled. Lower is better. A COT of 1.0 means the robot uses 1 joule of energy to move 1 newton of body weight 1 meter. The human body, with its evolved biomechanical optimization, achieves a COT of approximately 0.2 at comfortable walking speeds.
Historical robot COT numbers put the challenge in sharp perspective:
• ASIMO (Honda, circa 2011): COT approximately 3.23 — over 16 times less efficient than a human walker. Early humanoid robots were demonstrations of capability, not efficiency.
• Boston Dynamics Atlas (hydraulic): COT approximately 1.7 — still 8.5x worse than human, but reflecting meaningful improvement from fully actuated joints and real-time gait optimization.
• Duke Humanoid (research platform): COT approximately 0.77 — approaching 4x human efficiency, representing a breakthrough through passive dynamics research and compliant joint design.
• Agility Digit (target): COT below 0.5 — the threshold Agility Robotics has identified as commercially viable for warehouse and logistics applications.
The techniques driving COT improvement include passive dynamics (allowing limbs to swing naturally rather than actively driving every degree of freedom), spring-loaded joints that store and return elastic energy during each stride (mimicking the function of Achilles tendons), and optimized gait planning that selects the most efficient trajectory for a given speed and payload. Energy-aware motion planning algorithms can reduce walk-cycle power consumption by 20–30% compared to brute-force trajectory optimization alone.
Why COT Matters for Business
A robot with COT 0.5 walking the same distance as one with COT 2.0 uses four times less energy for locomotion. At 500 watts average total draw, that difference can mean 2 hours of runtime versus 8 hours — the difference between a robot that needs mid-shift battery swaps and one that completes a full industrial shift uninterrupted. The commercial calculus is not subtle.
The Compute Tax: AI Power Bill
Modern humanoid robots are not just mechanical systems — they are mobile data centers. The onboard compute required for real-time perception, navigation, manipulation planning, and natural language interaction imposes a significant and growing power cost that sits entirely outside the mechanical efficiency equation.
Consider the NVIDIA Jetson Thor, the current high-end reference platform for humanoid compute. It delivers 2,070 TOPS of AI compute and draws 40–130 watts depending on workload — representing 8–26% of a typical 500-watt walking budget. For context, the human brain performs vastly more complex perception and motor coordination tasks at just 20 watts. The efficiency gap between biological and silicon neural computation remains enormous.
Figure's Gen 2 robot takes this further, using dual RTX GPU modules for onboard inference — a configuration that prioritizes capability over power efficiency. This reflects a deliberate trade: accept higher power draw now in exchange for the richer perception and manipulation capabilities that make the robot commercially useful in demanding environments.
• Edge AI optimization: Techniques like quantization (reducing model weights from 32-bit to 8-bit or 4-bit), pruning, and knowledge distillation can reduce inference power by 2–4x with minimal accuracy loss. This is increasingly standard practice for production deployments.
• Heterogeneous compute: Running different workloads on different accelerators (CPU for control loops, NPU for vision, GPU for LLM inference) allows power gating of idle subsystems, reducing idle compute draw significantly between active tasks.
• Neuromorphic computing: IBM NorthPole — built on a 22nm process with 2,048 cores — demonstrates that fundamentally different computing architectures can achieve orders-of-magnitude better efficiency for specific AI workloads. Not yet production-ready for general robotics, but a clear long-term trajectory for the field.
AI-generated image
Making It Work Today: Operational Strategies
While the long-term solution to humanoid runtime is better batteries and more efficient compute, the near-term reality requires operational workarounds that allow productive deployment within current constraints. Several approaches are gaining traction:
• Hot-swappable battery packs: The Unitree H1 pioneered accessible battery swap design, allowing a depleted pack to be replaced in under two minutes without tools. A robot with 90 minutes of runtime can maintain near-continuous operation if a human or automated handler swaps packs on a regular schedule.
• Fleet rotation: Rather than trying to get 8 hours from one robot, operators deploy multiple robots in rotation — one working while another charges. This trades capital cost for operational continuity, a trade that makes sense when robot unit costs are still declining rapidly.
• Energy-aware motion planning: Software that optimizes trajectories for energy consumption rather than pure speed can reduce power draw by 20–30% on repetitive tasks. A robot restocking shelves can be routed to minimize total joint torque even if it takes 10% longer per trip.
• 48V fast charging infrastructure: Charging stations capable of 10–20 kW input can refill a 2.3 kWh pack in 7–14 minutes. Installing this infrastructure as part of facility buildout is becoming a standard requirement for humanoid deployment contracts.
• Battery buffering at stations: Some operators pre-charge spare packs in buffer storage, similar to power tool battery management in construction. The charging station becomes a small-scale energy buffer rather than a direct grid connection, smoothing demand peaks.
What Comes Next: The Solid-State Roadmap
The humanoid industry is watching one technology above all others: solid-state batteries. By replacing the liquid electrolyte in current lithium-ion cells with a solid ceramic or polymer material, solid-state designs promise dramatic improvements in both energy density and safety.
The roadmap, as the industry currently understands it:
• 2025: Early solid-state cells achieving approximately 250 Wh/kg at the cell level are entering pilot production. Toyota, QuantumScape, and Solid Power are among the leaders. These are not yet cost-competitive with lithium-ion at volume, but proving the chemistry works at scale.
• 2028–2030: Projected crossover point where solid-state cells reach cost parity and begin displacing lithium-ion in premium applications including robotics. This is when the first commercial humanoid robots with meaningfully extended runtime are likely to appear.
• 2035 target: Cell energy density of 500 Wh/kg — more than double current lithium-ion. At that density, a battery pack with the same mass as today Optimus Gen 2 unit would deliver approximately 4.6 kWh, enabling potential 8-hour runtimes under typical operating loads without any increase in robot weight.
The scale of demand the humanoid industry will create for batteries is staggering. Industry analysts project humanoid battery demand growing from approximately 0.05 GWh in 2025 to 74 GWh by 2035 — a 1,500x increase in a decade. That demand signal is already reshaping battery supply chain investment decisions globally.
Two complementary technologies are also being explored for specific use cases:
• Supercapacitors (EDLCs): Maxwell Technologies and others make electrochemical double-layer capacitors that can accept and deliver energy far faster than batteries. The application in robotics is peak burst suppression — using a supercapacitor buffer to handle 1–2 kW spikes during heavy lifting without stressing the battery pack. This can extend pack cycle life significantly by smoothing the current profile.
• Hydrogen fuel cells: Theoretically compelling — hydrogen has a gravimetric energy density of 33 kWh/kg, compared to 0.25 kWh/kg for lithium-ion. A fuel cell system could enable 8-plus hour runtimes at low system weight. The barrier is infrastructure: hydrogen refueling stations are rare, storage tanks add volume, and the system adds mechanical complexity. Fuel cells are unlikely to appear in general-purpose humanoid products before 2030 at the earliest.
The Industrial Stakes: Why Runtime Is a Business Metric
Runtime is not an abstract engineering specification. It maps directly to the unit economics of humanoid robot deployment, and those economics determine whether a particular application is commercially viable at all.
The global robot power management market was valued at approximately $640 million in 2024 and is projected to reach $1.68 billion by 2030, a compound annual growth rate exceeding 17%. That growth is driven almost entirely by the expectation that humanoid robots will begin commercial deployment at scale during this window.
The 8-hour shift threshold is the metric that operators consistently cite as the commercial viability marker. Here is why the math works out that way:
The Shift Station Calculus
Assume a robot-assisted assembly station running one 8-hour shift per day. If each robot lasts 4 hours, you need 2 robots per station to maintain continuous coverage — one working while one charges. If robots last 8 hours, you need 1 robot per station. With robots currently priced at $30,000–$200,000 depending on capability, the capital cost difference is significant — but the operational cost difference (scheduling, battery management, maintenance cycles) is often even larger.
Operators running multiple stations make this calculation across their entire deployment. A 100-station facility that can use 1 robot per station instead of 2 saves 100 robot units of capital plus the associated infrastructure and management overhead. That is not a marginal improvement — it is the difference between a business case that pencils out and one that does not.
This economic pressure explains why power architecture is receiving a disproportionate share of R&D investment relative to its apparent glamour. Perception algorithms and manipulation dexterity are the capabilities that make demos impressive. Battery runtime is the capability that makes deployments profitable.
Frequently Asked Questions
Why do humanoid robots only last 1–4 hours on a charge?
Walking draws 500-plus watts continuously, and most robots carry 2–2.3 kWh packs. At that ratio, 2–4 hours is the practical ceiling even at 100% efficiency. Onboard AI compute adds another 40–130 watts on top of the motor load. Add losses in the power electronics, battery degradation over cycles, and the reality that robots are not always running at minimum power, and real-world runtimes cluster toward the lower end of that range during active duty cycles. The fundamental constraint is energy density: lithium-ion cells cannot store enough energy per kilogram to do better given current robot power consumption profiles.
Why not just use a bigger battery?
Battery mass must sit in the torso for balance reasons — it cannot be distributed freely without destabilizing the robot gait. Heavier battery packs increase the load on every joint during motion, which increases power draw, which partially cancels the runtime gain from the added capacity. The standard engineering ceiling is approximately 1/8 of total robot body weight. For a 57 kg robot, that is about 7 kg of battery — roughly corresponding to the 2.3 kWh packs used today. A doubling of pack capacity would require a battery that violates this ratio, negatively impacting agility, joint wear, and payload capacity.
When will humanoid robots be able to work a full 8-hour shift?
The practical path to 8-hour shifts runs through two phases. Near term (2025–2028): hot-swappable packs and fleet rotation allow continuous coverage with current 2–4 hour runtimes. Medium term (2028–2032): solid-state batteries entering commercial production should deliver cell densities of 300–400 Wh/kg, enabling packs that fit within the 1/8 weight rule but hold 4–6 kWh. A true single-charge 8-hour runtime requires cell energy density roughly double what is commercially available today — achievable with solid-state chemistry, but not before the end of this decade.
Will solid-state batteries fix the runtime problem?
Solid-state batteries are the most promising technology on the roadmap, but they are not a complete solution on their own. Even at 500 Wh/kg cell density (the 2035 target), a robot still draws 500 watts walking. The improvement is that the same pack weight stores twice as much energy, doubling runtime from the energy-density side. The other half of the equation — reducing power consumption through more efficient motors, drives, and compute — must advance in parallel. The realistic scenario for 8-hour runtimes combines solid-state cells with continued improvements in motor efficiency, lighter structures, and more power-efficient AI inference chips.
How does robot energy efficiency compare to humans?
The comparison is humbling. A human walks at a Cost of Transport of approximately 0.2 and runs the most powerful perception-and-motor-control system on Earth at 20 watts. Current humanoid robots have COT values of 0.77–3.23 and run their AI at 40–130 watts. The mechanical efficiency gap is roughly 4–16x depending on the platform. The computational efficiency gap — measured in useful intelligence per watt — is far larger. The long-term trajectory for both gaps is improvement, but the human body remains the most efficient general-purpose locomotion and manipulation system ever built, and closing that gap is measured in decades, not years.
The Bottom Line: The power problem is not a footnote in humanoid robotics — it is the central constraint shaping every commercial decision in the industry. Robots that can walk and work for 2–4 hours before needing attention are useful demonstrations. Robots that can run an 8-hour shift without interruption are a manufacturing revolution. The gap between those two realities is measured in kilowatt-hours per kilogram, and the path from here to there runs through solid-state battery chemistry, GaN power electronics, neuromorphic computing, and energy-aware software that treats every joule as a scarce resource.