From Temperature-Limited to Battery-Limited Legged Locomotion

Raphaël Ségas, Balázs Opra, Léo Ghafari, Thomas StewartarXiv 2026 · arXiv

TL;DR

Legged robots often hit thermal limits and shut down long before their batteries are empty. By incorporating a reward based on predicted thermal steady-states into reinforcement learning, researchers enabled a Unitree Go2 quadruped to increase its endurance from 19 minutes to nearly 3 hours, effectively shifting the bottleneck from motor temperature to battery capacity.

What Problem Does This Paper Address?

In the world of legged robotics, agility is no longer the primary hurdle; endurance is. While modern controllers can navigate stairs and rough terrain with ease, they often do so at a high thermal cost. Electric motors generate heat—specifically Joule heating, which scales with the square of the torque. In the compact, poorly ventilated legs of a quadruped, this heat builds up faster than it can escape.

The result is a "thermal shutdown": the robot stops to protect its motors while the battery is still mostly full. For reinforcement learning (RL), this is a nightmare to solve. Thermal time constants (how long it takes a motor to heat up) are measured in tens of minutes, while RL training episodes last only seconds. This makes "credit assignment"—figuring out which specific movement caused the overheating—nearly impossible for standard algorithms.

Methods

The authors propose a "steady-state-aware" thermal reward. Instead of trying to teach the robot to monitor its current temperature, they teach it to care about where its temperature is headed.

  1. Lumped RC Thermal Model: They developed a 13-node thermal network (12 motors + 1 body node) with 22 parameters. Crucially, they "anchored" this model using electrical power balance. By measuring total bus power, they ensured the fitted parameters (like friction and resistance) remained physically realistic rather than just mathematically convenient.
  2. Predicting the Asymptote: The key innovation is calculating the thermal steady-state ($T_{ss}$). For any given gait (defined by its average torque and joint speed), there is a specific temperature at which the motor will eventually stabilize. While the actual temperature moves slowly, this predicted asymptote responds instantly to changes in the gait.
  3. Soft-Max Reward: Overheating is a distribution problem—it only takes one motor to fail for the whole robot to stop. The RL reward penalizes a "soft maximum" of the predicted asymptotes across all 12 joints. This incentivizes the policy to not just minimize energy, but to "spread the heat" across different joints.
Comparison of baseline (temperature-limited) vs. the proposed thermal policy (battery-limited).

Datasets

The study utilized a comprehensive dataset collected from a Unitree Go2:

  • Training Data: Two 4-hour runs involving sweeps of velocities, payloads (0–5 kg), and various static stances to identify thermal parameters.
  • Validation/Test Data: 21 held-out runs totaling 39 hours of data across two different robots, various controllers, and payloads up to 10 kg.
  • Public Release: The authors have open-sourced the "Unitree Go2 Motor Thermal Dataset," providing motor temperatures, torques, and electrical statistics for future research.

Key Insights and Ideas

  • Collapsing the Time Scale: By rewarding the asymptote rather than the state, the authors collapsed the credit-assignment horizon from minutes down to a single gait cycle. This allows standard RL tools like PPO to "see" the thermal consequences of their actions immediately.
  • Posture as a Heat Sink: The policy learned that posture matters. To stay cool, the robot adopts a more upright stance with straighter legs, which reduces the torque required by the calf motors—the most frequent victims of overheating.
  • Distribution over Summation: Traditional energy minimization (minimizing the sum of squared torques) doesn't prevent individual joints from overheating. The p-norm (soft-max) approach specifically targets the "weakest link" in the thermal chain.

Key Results

The performance gains on the physical Unitree Go2 were dramatic:

  • Endurance: With a 5 kg payload at 1 m/s, the baseline policy overheated in just 19 minutes (1.1 km). The thermal policy ran until the battery died at 174 minutes (10.0 km), maintaining a 14°C safety margin.
  • Efficiency: The thermal policy was significantly more economical, drawing 142 W compared to the baseline's 244 W—a 43% reduction in the electrical cost of transport.
  • Versatility: The policy maintained its ability to navigate difficult terrain. On a factory circuit featuring steep stairs and ramps, the thermal policy completed 47 laps (11.2 km) on a single charge, whereas the baseline overheated after 25 laps.
The hottest joint temperature on the factory circuit: the baseline (purple) hits shutdown quickly, while the thermal policy (green) stays well below the limit.

Limitations and Future Work

While highly effective, the authors noted a small sim-to-real gap, particularly at higher payloads where predicted asymptotes were slightly lower than measured ones.

Future directions include:

  • Closing the sim-to-real gap for heavy loads.
  • Developing an explicit electrical power reward to further maximize range.
  • Implementing a "thermal governor" that can dynamically adjust the robot's commanded velocity based on real-time temperature predictions to ensure it never hits a shutdown limit, regardless of the environment.

Summary updated

Abstract

On legged robots, sustained locomotion often ends in thermal shutdown rather than battery depletion: the heat of walking concentrates in a few joint motors, which overheat long before the battery is empty. We train thermally aware locomotion policies by reinforcement learning without ever observing temperature, and demonstrate this on a Unitree Go2 quadruped. We first identify a coupled 13-node lumped RC thermal network, fitted on two runs of a single robot and calibrated against the measured electrical power, that predicts all twelve motor temperatures with a mean RMSE below 1 $^\circ$C across different robot instances, gaits, payloads, and speeds. Under a sustained gait, the temperatures the motors will settle at follow directly from the gait-averaged squared torques and joint speeds. Our core contribution is a reward that drives down the hottest predicted asymptotes: thermal time constants exceed the credit-assignment horizon of reinforcement learning by orders of magnitude, but predicted asymptotes respond to torque and joint speed within a single gait cycle, so the policy learns to lower and spread heat with no temperature input. On the robot, the policy turns sustained locomotion from temperature-limited to battery-limited: with a 5 kg payload at 1 m/s, a learned baseline reaches thermal shutdown after 19 minutes (1.1 km), whereas our policy runs the battery down after 174 minutes (10.0 km) with its hottest motor 14 $^\circ$C below the shutdown threshold in a 26 $^\circ$C room; on an unloaded factory circuit with stairs and ramps, the baseline overheats after 25 laps (6.2 km), while our policy completes 47 (11.2 km) on one battery with 22 $^\circ$C of margin.
BibTeX
@misc{segas2026from,
  author = {Raphaël Ségas and Balázs Opra and Léo Ghafari and Thomas Stewart},
  title = {{From Temperature-Limited to Battery-Limited Legged Locomotion}},
  year = {2026},
  eprint = {2611.22231},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22231},
}
Figure