Research map / Label-efficient and robust learning / Test-time adaptation

Efficient Test-Time Adaptation For Robot Policies

NeurIPS 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Pseudo-label test-time adaptation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TEMPO, a reward-free test-time adaptation method for RL locomotion policies that updates only the actor against a frozen distributional critic, using an entropy regularizer and a score-based importance-sampled memory.

Problem

Pretrained RL locomotion policies degrade under deployment-time distribution shifts (payloads, rough or slippery terrain), and offline retraining is infeasible under on-device latency and memory limits.

Approach

Adapt only the actor by maximizing a frozen pretrained critic's value on a small memory of observations, add an entropy term on the critic's distributional output to avoid overconfident extrapolation, and replace FIFO memory with a score-based buffer that keeps low-value "hard" observations and samples them via importance sampling.

Key results

• TEMPO yields up to 80% improvement over no adaptation in a single-device simulation setting across Go1, Go2, and G1 under four shifts.
• On a physical Unitree Go2 with calf payload, TEMPO raises mean real-world reward from 89.8 to 98.7 versus the non-adapted baseline.
• Under mixed corruptions, TEMPO averages 14.09 environment reward vs 11.18 for no adaptation and 12.48 for QMAX.

Surprising findings

• Test-time training with a KL regularizer monotonically reduces return as the regularization coefficient grows, so the best TTT setting reduces to plain QMAX.

Limitations

• Relies on a pretrained critic that is reasonably informative under shift; a miscalibrated critic can degrade adaptation.
• Adds a small memory and compute overhead at inference time.
• Real-world validation is limited to a single calf-payload shift on one robot.

Experimental setup

Datasets: MuJoCo Playground GO1-JOYSTICK MuJoCo Playground GO2-JOYSTICK MuJoCo Playground G1-JOYSTICK

Benchmarks: MuJoCo Playground GO1-JOYSTICK MuJoCo Playground GO2-JOYSTICK MuJoCo Playground G1-JOYSTICK Unitree Go2 real-world calf-mass shift

Metrics: episodic return average forward linear velocity angular velocity tracking reward real-world velocity tracking reward

Baselines: No Adapt QMAX Tent OPEX Test-Time Training RMA

Builds on: Tent Fast-TD3 Rapid Motor Adaptation prioritized experience replay MuJoCo Playground

Closest papers by meaning

• Test-Time Adaptation for Online Vision-Language Navigation with Feedback-based Reinforcement Learning (ICML 2025 · similarity 0.89)
• GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Learning to Reuse Policies in State Evolvable Environments (ICML 2025 · similarity 0.88)
• Test-time Adapted Reinforcement Learning with Action Entropy Regularization (ICML 2025 · similarity 0.88)
• PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.87)
• Test-time Risk Adaptation with Mixture of Agents (NeurIPS 2026 · similarity 0.87)