Event
RoboLongR: Causally Consistent Reward Learning for Long-Horizon Manipulation
1 reports · 1 sources
Zixuan Chen and co-authors propose RoboLongR, an offline-to-online reward learning framework: through causal consistency, it aligns reward progress with the preconditions supported by the execution history, guiding the policy to transition from behavior cloning to online reinforcement learning. The method is pretrained on demonstrations from 88 RLBench tasks and Open-X data; on 18 long-horizon tasks that did not participate in reward pretraining, it achieves an average success rate of 31.8%, 9.2 percentage points higher than the strongest comparison baseline; on four real-robot tasks, the success rate is 57.5%, 32.5 percentage points higher than the comparison methods.
Report timeline
- RoboLongR: Causally consistent rewards guide robots to progressively master long-horizon manipulation
RoboLongR proposes an offline-to-online reward learning framework that uses causal consistency to align reward progress with the prerequisites supported by execution history, guiding policies from behavior cloning to online reinforcement learning. Pretrained on 88 RLBench tasks and Open-X data, the method achieves an average success rate of 31.8% on 18 long-horizon tasks excluded from reward pretraining, 9.2 percentage points above the strongest comparison baseline; on four real-robot tasks it achieves a 57.5% success rate, 32.5 percentage points above the comparison method.
arXiv cs.RO ·