RACE: Calibrated Cumulative Scoring for Rollout-Free Robot Policy Checkpoint Selection
RACE: Calibrated Cumulative Scoring for Rollout-Free Robot Policy Checkpoint Selection
This paper presents RACE, a new metric for selecting robot policy checkpoints without costly real-world trials. Unlike averaged validation loss, which can misrank checkpoints, RACE calibrates action deviations against state-conditional thresholds derived from successful trajectories. By aggregating scores via a running product, it reduces the impact of early deviations on subsequent predictions. Experiments in simulation and on real robots show that RACE outperforms baselines across various policy backbones and tasks, with rankings aligning well with deployment success.
The authors introduce RACE, a rollout-free metric that calibrates action deviations using state-conditional thresholds derived from successful trajectories to better predict deployment performance than averaged validation loss.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Source：arXiv Robotics — research abstracts · arxiv.org