HiRE improves robot policy finetuning by editing rewards hindsight
HiRE: Hindsight Reward Editing for Policy Finetuning
Robotics
Summary
Training robots to perform tasks requires fine-tuning their decision-making based on rewards, but designing good rewards is hard and can be misleading. The authors propose a new method called HiRE that looks back at completed tasks to find mistakes where the robot was wrongly rewarded or penalized. By adjusting these rewards after the fact, HiRE helps the robot learn more effectively and improves its performance. This approach works with different learning algorithms and representations without extra training.
What this means in practice
- •For robotics engineers: Improve adaptation and training speed of robot control policies in new environments by correcting reward signals retrospectively.
- •For autonomous system developers: Enhance the stability and performance of autonomous decision systems by filtering misleading rewards from learned behavior data.
Authors
Haoyi Niu, Zhengtao Han, Yufeng Ji, Zhongyu Li, Koushil Sreenath
Abstract
Pre-trained robot policies always require finetuning to adapt to specific environments. Reinforcement Learning (RL) offers high performance potential because it improves action optimality rather than simply mimicking data. However, such potential depends heavily on reward quality. Sparse rewards lack process feedback, human-designed rewards are costly and biased, and semantic rewards from foundation representations are often not control-centric. We propose Hindsight Reward Editing (HiRE), a training-free framework to break this reward bottleneck. HiRE bridges the broad knowledge of foundation representation models with physical control awareness, by contrasting successful and failed trajectories in hindsight. It calibrates foundation representation models by identifying "trap states" that are predicted as high-rewarding states yet eventually result in failure, and vice versa. HiRE explicitly penalizes these traps while boosting rewards for critical successful states. This approach can be flexibly compatible with any foundation representations and RL algorithms. Experiments show that HiRE consistently outperforms other reward recipes by delivering dense, control-aware feedback that prevents value function collapse and reward hacking, thereby achieving superior sample efficiency, stable policy updates, and higher performance ceilings, e.g., at least 3x performance of the base policies. Qualitative results are at https://hire-project.github.io .