← Paper Digest

From Temperature-Limited to Battery-Limited Legged Locomotion

Raphaël Ségas, Balázs Opra, Léo Ghafari, Thomas Stewart

November 2026

TL;DR A reward that penalizes the motor temperatures a gait will eventually settle at, predicted by an identified thermal model, teaches a Unitree Go2 locomotion policy to spread heat without ever observing temperature. Carrying 5 kg at 1 m/s it walks for 174 minutes until the battery is empty, where a learned baseline overheats after 19.

What they do

On legged robots, long runs often end when one joint motor overheats, not when the battery runs out. The authors make an RL locomotion policy for the Unitree Go2 thermally aware without giving it any temperature input. Motor temperatures change over tens of minutes, far longer than an RL episode, but the temperatures a sustained gait will settle at follow from that gait’s averaged squared torques and joint speeds, which change within one gait cycle.

How it works

  • Thermal model. A 22-parameter, 13-node lumped RC network (twelve motors plus a hidden body node), fitted on two four-hour runs of one robot. Quiescent power, the Joule heating coefficient and the resistance temperature coefficient are first fixed from the measured electrical power balance, which keeps the fitted parameters physical.
  • Asymptote reward. Setting the temperature derivative to zero gives a 13x13 linear solve for the steady-state temperatures, cheap enough for every simulation step. The reward penalizes a soft maximum of these predicted asymptotes, so it targets the hottest joint rather than total energy.
  • Training. The reward is added to a DreamWaQ re-implementation trained with PPO in Isaac Lab on stairs and slopes, with 0-5 kg payload randomization.

Results

  • The thermal model predicts all twelve motor temperatures with RMSE below 1 °C on 21 held-out runs, including a second robot and payloads up to 10 kg.
  • In simulation with 5 kg, the hottest predicted asymptote is 72 °C, against 80 °C for a baseline with a squared-torque penalty and 131 °C for the plain baseline. Pooled survival on hard terrains stays at or above 99.7%.
  • On the robot with 5 kg at 1 m/s: 174 min (10.0 km) until the battery is empty, against thermal shutdown after 19 min (1.1 km), drawing 142 W instead of 244 W.
  • Unloaded factory circuit with 19.5 cm stairs: 47 laps (11.2 km) against 25 laps (6.2 km), hottest joint peaking at 64 °C.

Limitations

The sim-to-real gap grows with payload: measured gaits give asymptotes 4.6 °C above simulation at 5 kg. In simulation the policy has 18-19% more velocity-tracking error on stairs. Baseline and policy hardware runs were on different days, with the baseline on the cooler one. Only shown on the Go2 with a blind policy.

Figures

From Figure 1: Sustained 1 m/s walking with a 5 kg payload. The baseline reaches thermal shutdown after 19 min (top);...
From Figure 1: Sustained 1 m/s walking with a 5 kg payload. The baseline reaches thermal shutdown after 19 min (top); the thermal policy drains the battery after 174 min with 14 °C of headroom (bottom). Grey: the 12 measured motor temperatures; shaded: range of the thermal model; dots: predicted asymptotes (blue hip, orange thigh, green calf).
From Figure 2: The two training runs used to fit the thermal model, protocol A (no payload) and protocol B (5 kg payl...
From Figure 2: The two training runs used to fit the thermal model, protocol A (no payload) and protocol B (5 kg payload), mixing walking, stand-lock and damp segments. The twelve motor temperatures are logged every two seconds; their maximum stays under the Go2 shutdown threshold of about 86 °C.
From Figure 4: Hottest predicted motor asymptote versus payload at 1 m/s in simulation, for three thermal-reward weig...
From Figure 4: Hottest predicted motor asymptote versus payload at 1 m/s in simulation, for three thermal-reward weights and the strengthened baseline (means and s.d. over five seeds and five checkpoints). Crosses: the two deployed runs of the policy with weight -0.2.
From Figure 7: Factory circuit, no payload: hottest joint of the baseline (purple) and the thermal policy (green); me...
From Figure 7: Factory circuit, no payload: hottest joint of the baseline (purple) and the thermal policy (green); measured (solid) and predicted (dashed).
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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.