---

Abstract #

Power system design represents the single greatest constraint on humanoid robot autonomy. Current-generation humanoid platforms achieve only two to four hours of continuous operation, with battery mass consuming fifteen to twenty-five percent of total system weight and peak actuator demands creating discharge profiles fundamentally different from those in electric vehicles or consumer electronics. This article presents a complete power system architecture for open-source humanoid robots, spanning cell chemistry selection, pack topology design, battery management system implementation, power distribution network engineering, energy harvesting through regenerative braking, and runtime optimization via energy-aware task scheduling. We analyze the trade-offs between lithium-ion nickel manganese cobalt cells that dominate current platforms and emerging solid-state chemistries projected to reach 75 GWh of humanoid robot demand by 2035. The article introduces a hierarchical power budget framework that allocates energy across locomotion, manipulation, computation, and sensing subsystems while maintaining reserves for safety-critical functions. Through first-principles analysis of discharge profiles during bipedal walking, stair climbing, and manipulation tasks, we establish design guidelines for pack sizing, bus voltage selection, and dynamic power management that enable open-hardware designers to engineer reliable power systems without proprietary simulation dependencies.

1. Introduction #

In the previous article, we examined thermal management for humanoid robots, establishing how heat dissipation strategies, actuator cooling architectures, and operating temperature envelopes ensure reliable continuous operation under sustained thermal loads [1^[2]]. Thermal constraints and power system design are deeply coupled: battery cells exhibit capacity fade and internal resistance increases at elevated temperatures, while thermal management systems themselves consume power that must be budgeted from the same energy store they protect.

The power challenge facing humanoid robotics is stark. According to TrendForce analysis published in January 2026, most current humanoid robots achieve a runtime of only two to four hours on a single charge, with increasing that operating time depending on either hot-swappable batteries or higher-energy-density cell chemistries [2^[3]]. This limitation is not merely an inconvenience but a fundamental barrier to commercial deployment: warehouse logistics, eldercare assistance, and manufacturing support all require eight-hour or longer operational windows to justify the capital investment in humanoid platforms.

The problem is compounded by the unique discharge profile of bipedal robots. Unlike electric vehicles that operate primarily at steady-state cruising with occasional acceleration peaks, humanoid robots experience continuous high-frequency power transients as leg actuators cycle through stance and swing phases at every step [3^[4]]. Hip and knee joints demand peak currents during push-off that can exceed sustained averages by three to five times, creating pulsed discharge patterns that stress cell chemistry, complicate state-of-charge estimation, and accelerate capacity degradation.

As Interact Analysis noted in their January 2026 assessment, robot architecture is evolving too rapidly to determine which battery technology will prove dominant, with ongoing changes in joint actuation designs, form factors, thermal management solutions, and edge AI power consumption leaving future space allocation and discharge requirements uncertain [4^[5]]. This article addresses that uncertainty by establishing first-principles design methodology rather than prescribing specific cell vendors, enabling open-source humanoid robot builders to adapt their power systems as both robot architectures and battery technologies mature.

2. Cell Chemistry and Pack Architecture #

The foundation of any humanoid robot power system is the electrochemical cell. Selecting the appropriate chemistry requires balancing energy density against power density, cycle life, thermal stability, and cost within the unique constraints of an anthropomorphic form factor.

2.1 Lithium-Ion Chemistry Landscape #

Current humanoid robots overwhelmingly use lithium-ion cells, but the specific cathode chemistry varies significantly across platforms. Nickel manganese cobalt (NMC) cells in the 811 formulation deliver gravimetric energy densities of 250-300 Wh/kg at the cell level, making them the default choice for weight-sensitive bipedal platforms. Lithium iron phosphate (LFP) cells sacrifice energy density at 160-190 Wh/kg but offer superior thermal stability and cycle life exceeding 3000 full cycles, attracting designers who prioritize longevity over runtime [5^[6]].

The choice between NMC and LFP is not merely technical but economic. As the MANLY Battery 2026 program guide details, humanoid robot battery packs remain highly customized, with duty cycles that are peak-heavy and uptime targets that force engineering teams to size energy and power budgets from real operational traces rather than theoretical calculations [5^[6]]. For an open-source platform where operational profiles may vary dramatically between builder applications, this customization requirement argues for a modular pack architecture that can accommodate different cell chemistries without redesigning the structural integration.

flowchart TD
    A[Cell Chemistry Selection] --> B{Primary Requirement}
    B -->|Maximum Runtime| C[NMC 811]
    B -->|Maximum Cycle Life| D[LFP]
    B -->|Maximum Safety| E[LFP / Solid-State]
    B -->|Maximum Power Density| F[NMC with Supercap Buffer]
    C --> G[250-300 Wh/kg]
    D --> H[160-190 Wh/kg]
    E --> I[160-400 Wh/kg]
    F --> J[Hybrid Energy/Power]
    G --> K[Pack Design: Series-Parallel Config]
    H --> K
    I --> K
    J --> K
    K --> L[BMS Integration]
    L --> M[Power Distribution Network]

2.2 Solid-State Battery Horizon #

The most significant near-term disruption to humanoid robot power systems is the emergence of solid-state batteries. TrendForce projects that demand for solid-state batteries in humanoid robots may surpass 74 GWh by 2035, representing more than a thousand-fold increase from 2026 levels [2^[3]]. Solid-state cells replace the liquid organic electrolyte with a solid ceramic or polymer conductor, eliminating the primary source of thermal runaway risk while enabling theoretical energy densities exceeding 400 Wh/kg.

However, the transition is not immediate. As Figure AI emphasized during their F.03 platform launch, off-the-shelf electric vehicle batteries cannot be directly reused in humanoid robots, and customized solid-state batteries must wait until robot system architectures stabilize [4^[5]]. For open-source builders in 2026, this means designing pack enclosures and bus architectures that can accommodate solid-state cells when they become commercially viable while using proven lithium-ion chemistry today.

Nature Communications research on solid-state lithium-sulfur batteries demonstrates the promise of significantly enhanced energy density compared to current lithium-ion technology, with particular relevance for weight-constrained mobile platforms [6^[7]]. The practical implication for humanoid robot designers is to standardize on cell form factors and voltage ranges that overlap between current NMC cells and anticipated solid-state equivalents, enabling chemistry-agnostic pack upgrades.

2.3 Pack Topology for Humanoid Form Factors #

A humanoid robot presents unique geometric constraints for battery pack placement. The pack must maintain a low center of gravity to support bipedal balance while distributing mass symmetrically about the sagittal plane. Most current platforms locate the primary battery in the torso cavity, but this creates thermal coupling with compute modules and trunk actuators.

Parameter	Torso-Mounted	Distributed	Hip-Belt
CoG Height	Medium (450mm)	Low (350mm)	Low (320mm)
Balance Impact	Moderate	Minimal	Minimal
Thermal Isolation	Poor	Good	Good
Wiring Complexity	Low	High	Medium
Swap Access	Rear Panel	Multiple Points	Waist Access
Structural Load	Concentrated	Distributed	Pelvic Frame
Typical Capacity	1.5-2.5 kWh	1.0-2.0 kWh	1.0-1.8 kWh

The distributed approach divides the total energy storage across multiple smaller packs located in the thighs, torso, and potentially the upper arms. While this improves mass distribution and thermal management, it increases wiring complexity and requires sophisticated BMS coordination across physically separated cell groups. For open-source designs targeting simplicity, a torso-mounted pack with a well-engineered thermal barrier between cells and compute hardware represents the pragmatic starting point, with provisions for future migration to distributed architectures.

3. Battery Management and Power Distribution #

The battery management system (BMS) serves as the intelligence layer between raw electrochemical energy and the robot’s operational demands. In humanoid robots, BMS requirements exceed those of typical consumer or automotive applications due to the combination of high peak-to-average power ratios, rapid load transients, and the safety implications of a heavy mobile platform operating near humans.

3.1 State Estimation Under Dynamic Loads #

Accurate state-of-charge (SOC) estimation is particularly challenging in humanoid robots because standard coulomb counting methods accumulate errors rapidly under pulsed discharge conditions. Recent work by Souza et al. at MDPI demonstrates online SOC estimation for lithium-ion batteries using multilayer perceptron neural networks applied to an instrumented robot, achieving estimation accuracy suitable for real-time operational decisions [7^[8]]. Their approach processes voltage, current, and temperature measurements through a trained neural network that captures the nonlinear relationship between observable electrical quantities and internal electrochemical state.

Zhou et al. in Scientific Reports present a complementary approach using extended Kalman filtering for real-time SOC estimation in autonomous charging robots, demonstrating that model-based state estimation can maintain accuracy even during the rapid charge-discharge cycles characteristic of mobile platforms [3^[4]]. For open-source implementations, the extended Kalman filter offers a practical balance between estimation accuracy and computational cost, requiring only a parametric equivalent circuit model rather than the large training datasets needed for neural network approaches.

flowchart LR
    subgraph Sensors
        V[Voltage Sensor]
        I[Current Sensor]
        T[Temperature Sensor]
    end
    subgraph BMS_Core
        CC[Coulomb Counter]
        EKF[Extended Kalman Filter]
        OCV[OCV Lookup Table]
        ML[ML Estimator]
    end
    subgraph Outputs
        SOC[State of Charge]
        SOH[State of Health]
        RTE[Runtime Estimate]
        SAFE[Safety Flags]
    end
    V --> EKF
    V --> OCV
    I --> CC
    I --> EKF
    T --> EKF
    T --> SAFE
    CC --> SOC
    EKF --> SOC
    OCV --> EKF
    ML --> SOH
    SOC --> RTE
    SOH --> RTE

A comprehensive review by ScienceDirect on deep learning applications in battery management systems highlights the transformative potential of neural network approaches for SOC estimation, state-of-health prediction, remaining useful life forecasting, fault detection, and energy optimization [8^[9]]. While full deep learning BMS implementations demand significant computational resources, embedded TinyML models can achieve SOC estimation with model sizes as small as 3.4 kilobytes and inference latencies under 2 milliseconds, making real-time neural BMS feasible even on the microcontrollers typical of open-source robot designs [9].

3.2 Power Distribution Architecture #

The power distribution network must deliver stable voltage rails to diverse subsystems with dramatically different power profiles. A typical humanoid robot requires at minimum three distinct power domains: a high-voltage bus (typically 48V) for leg and arm actuators, a regulated 12V rail for compute modules and sensors, and a 5V rail for communication interfaces and low-power peripherals.

Power Domain	Voltage	Peak Current	Duty Cycle	Priority
Leg Actuators	48V	40A	Continuous	Critical
Arm Actuators	48V	20A	Intermittent	High
Hand Actuators	24V	8A	Burst	Medium
Main Compute	12V	10A	Continuous	Critical
Vision/Sensors	12V	5A	Continuous	High
Communication	5V	2A	Continuous	Critical
LED/HRI	5V	1A	Intermittent	Low
Safety Systems	12V (backed)	3A	Always-On	Emergency

The safety-critical power domain deserves special attention. Emergency stop circuits, fall detection systems, and communication links must remain powered even during a main battery fault or deliberate shutdown. This requires either an independent safety battery or a supercapacitor bank that can sustain safety systems for the minimum time needed to execute a controlled shutdown sequence, typically 10-30 seconds for a humanoid robot to transition from walking to a stable resting pose.

3.3 Bus Voltage Selection #

The choice of main bus voltage involves trade-offs between actuator efficiency, cable mass, safety, and component availability. A 48V nominal bus has emerged as the de facto standard for humanoid robots, as it sits below the 60V DC safety threshold defined in IEC 61140 while providing sufficient voltage headroom for high-torque BLDC motor controllers. At 48V, a 2 kWh pack delivering 1 kW continuous power draws approximately 21A, allowing the use of moderate-gauge wiring that remains flexible enough to route through articulated joints.

Higher bus voltages of 72V or 96V would reduce current and thus cable mass, but exceed low-voltage safety thresholds and require additional isolation measures that increase system complexity. For open-source platforms prioritizing accessibility, the 48V standard provides the optimal combination of performance, safety, and component ecosystem maturity.

4. Energy Harvesting and Runtime Optimization #

Extending operational runtime beyond what raw battery capacity provides requires both recovering wasted energy and spending stored energy more intelligently. For humanoid robots, the two primary strategies are regenerative braking during locomotion and energy-aware computational scheduling.

4.1 Regenerative Braking in Bipedal Locomotion #

During each step cycle, a bipedal robot’s leg actuators alternate between powering phases, where motors drive joint rotation, and braking phases, where motors decelerate limb segments against gravitational and inertial loads. During braking, the BLDC motors function as generators, producing back-EMF that can be captured and returned to the battery through appropriately designed motor controllers.

Research on regenerative braking for electric vehicles demonstrates that energy recovery during deceleration can reduce total energy consumption by up to 30 percent [10^[10]]. In humanoid robots, the regenerative opportunity is somewhat smaller because gait cycles involve shorter deceleration windows than vehicle braking events, but the high frequency of recovery opportunities, two per step per active joint, accumulates meaningfully over sustained walking.

The practical recovery efficiency depends on motor controller topology. Field-oriented control drives with active rectification can achieve regenerative efficiencies of 60-80 percent of the theoretical kinetic energy available during limb deceleration. For a 70 kg humanoid walking at 1.2 m/s, the continuous regenerative power recovery from knee and hip joints combined typically reaches 15-25W, representing a 5-8 percent effective runtime extension during sustained locomotion. While modest in absolute terms, this recovery is essentially free once the motor controller hardware supports bidirectional current flow, which most modern FOC drives already do.

Advanced regenerative braking strategies employing fuzzy control and particle swarm optimization can further improve energy recovery by adapting braking torque distribution in real-time based on ground contact conditions and gait phase [11^[11]]. For humanoid robots, the analogous optimization adjusts regenerative torque profiles across hip, knee, and ankle joints to maximize total energy recovery while maintaining smooth gait transitions.

flowchart TD
    subgraph Gait_Cycle["Gait Cycle Energy Flow"]
        SW[Swing Phase] -->|Motor Power| JNT[Joint Actuator]
        JNT -->|Braking Phase| GEN[Generator Mode]
        GEN -->|Back-EMF| RECT[Active Rectifier]
        RECT -->|DC Return| BUS[48V Bus]
        BUS -->|Charge| BATT[Battery Pack]
        BATT -->|Discharge| INV[Motor Inverter]
        INV -->|Drive Current| JNT
    end
    subgraph Energy_Budget["Runtime Impact"]
        BASE[Base Runtime: 2.5h] --> REG[Regen Recovery: +8%]
        REG --> SCHED[Task Scheduling: +12%]
        SCHED --> TOTAL[Effective Runtime: 3.0h]
    end

4.2 Energy-Aware Task Scheduling #

Beyond energy recovery, significant runtime gains come from spending energy more intelligently. Energy-aware task scheduling treats the battery as a depletable resource and allocates power budgets across subsystems based on task priority, predicted duration, and remaining charge.

The core principle is dynamic power gating: subsystems that are not actively contributing to the current task receive reduced power or enter sleep states. During navigation-only phases, manipulation actuators can be powered down entirely, saving 50-100W. During stationary manipulation tasks, locomotion actuators enter a holding-torque mode that consumes 20-30 percent of walking power. Vision processing can switch between high-resolution object detection during manipulation and lower-resolution obstacle avoidance during transit, reducing compute power by 40-60 percent.

Operating Mode	Locomotion	Manipulation	Compute	Sensing	Total Power	Runtime (2kWh)
Full Active	350W	150W	75W	25W	600W	3.3h
Walk Only	350W	10W (idle)	45W	25W	430W	4.7h
Manipulate Only	30W (hold)	150W	75W	25W	280W	7.1h
Standby Aware	15W (lock)	5W (off)	30W	15W	65W	30.7h
Emergency Safe	5W (brake)	0W	10W	5W	20W	100h

This power budget framework reveals that the difference between full-active and intelligent scheduling can double effective runtime. The key implementation challenge is predicting task sequences accurately enough to pre-emptively transition subsystems between power states without introducing latency that degrades task performance.

4.3 Charging Strategy and Operational Continuity #

For continuous deployment scenarios, the charging strategy becomes as important as the power system design itself. Three approaches have emerged in the humanoid robotics ecosystem. Figure AI’s F.03 platform implements inductive charging through the robot’s feet, enabling autonomous dock-and-charge behavior analogous to consumer robot vacuums [12^[12]]. UBTECH’s Walker S2 demonstrates hot-swappable battery replacement where the robot autonomously exchanges its own depleted pack for a charged one, enabling theoretically continuous operation [13^[13]]. Traditional conductive charging through a docking connector remains the simplest and most efficient option for open-source implementations.

For an open-source platform, the pragmatic recommendation is to implement conductive docking as the primary charging method with a standardized pack form factor that enables manual hot-swap as a secondary option. Inductive charging through feet is elegant but requires custom coil integration and accepts efficiency losses of 10-15 percent compared to conductive connections. The hot-swap approach demands significant mechanical engineering for the latching mechanism and electrical design for live-disconnect handling but delivers the ultimate operational continuity.

5. System Integration and Design Guidelines #

Translating component-level analysis into a complete power system requires integrating cell chemistry selection, pack architecture, BMS implementation, power distribution, and runtime optimization into a coherent design that respects the constraints identified throughout this series.

5.1 Power Budget Methodology #

The design process begins with a comprehensive power budget derived from the actuator torque requirements established in Article 4 of this series and the sensing loads characterized in Article 7. For each anticipated operational mode, the designer must estimate both average and peak power consumption, then size the battery pack to deliver the required runtime at average power while supporting peak demands without excessive voltage sag.

A practical sizing formula for open-source builders targets a pack energy of 1.5 times the product of average power consumption and desired runtime, providing margin for capacity fade over the pack’s service life and for unanticipated peak loads. For a humanoid robot consuming 400W average power with a four-hour runtime target, this yields a minimum pack size of 2.4 kWh, which at current NMC 811 energy densities corresponds to approximately 8-10 kg of cells plus 2-3 kg of pack structure, BMS hardware, and interconnects.

5.2 Open-Source BMS Implementation #

Several open-source BMS platforms have matured sufficiently for humanoid robot applications. The critical requirements are support for the target cell count in series, which is typically 12-14 cells for a 48V nominal pack, active cell balancing during both charge and discharge, configurable protection thresholds for over-current, under-voltage, and over-temperature, and a communication interface, typically CAN bus, that integrates with the robot’s real-time control network.

The BMS must implement at minimum three protection layers: cell-level voltage monitoring with disconnect capability, pack-level current limiting with configurable thresholds for sustained and peak currents, and thermal monitoring with both cell-surface and ambient temperature inputs. These protection layers operate independently so that a fault in one monitoring path does not compromise the remaining safety functions.

5.3 Integration with Thermal Management #

The coupling between power systems and thermal management, analyzed in the previous article, requires coordinated design. Battery cells must remain within their safe operating temperature window, typically 15-45 degrees Celsius for lithium-ion chemistries, even when adjacent actuators generate significant waste heat. The thermal zoning framework established in Article 15 should include the battery compartment as a dedicated thermal zone with independent cooling capacity sized for both internal cell heating during high-rate discharge and conducted heat from neighboring actuator zones.

During fast charging, cells generate additional internal heat from resistive losses that can double the thermal load compared to normal discharge. The thermal management system must be designed for this worst-case condition, not merely for operational discharge loads.

6. Conclusion #

Power system design for autonomous humanoid robots operates at the intersection of electrochemistry, power electronics, control theory, and mechanical integration. The analysis presented in this article establishes several key principles for open-source humanoid robot builders.

First, cell chemistry selection should prioritize the specific discharge profile of bipedal locomotion rather than defaulting to the highest energy density option. The pulsed, high-frequency discharge patterns of walking create different aging mechanisms than the steady-state profiles for which most cell datasheets are optimized. NMC 811 cells currently offer the best energy-to-weight ratio for humanoid applications, but LFP chemistry deserves consideration for platforms where cycle life and thermal robustness outweigh runtime optimization.

Second, the battery management system must go beyond basic protection to provide accurate state estimation under dynamic loads. Extended Kalman filtering offers a practical starting point for open-source implementations, with TinyML approaches becoming feasible as training datasets from operational robots accumulate. Accurate SOC estimation directly impacts operational reliability by preventing unexpected shutdowns during critical tasks.

Third, runtime optimization through energy-aware task scheduling can effectively double operational endurance without any change to battery hardware. The power budget framework presented here, allocating energy across locomotion, manipulation, computation, and sensing based on current task demands, represents perhaps the highest-leverage engineering opportunity in humanoid robot power systems.

Looking forward, solid-state batteries promise to transform the energy density constraints that currently limit humanoid robot autonomy. With TrendForce projecting 75 GWh of solid-state battery demand for humanoid robots by 2035, open-source designers should standardize on pack form factors and bus architectures that accommodate future chemistry upgrades [2^[3]]. The next article in this series will examine communication protocols, including ROS 2, EtherCAT, and real-time networking, that connect the power management system to all other subsystems in a coherent, responsive humanoid robot architecture.