A robot may keep moving while making no progress. Event2Reward, introduced on October 29, learns to distinguish stalls, setbacks and recoveries. The estimated progress helps shape rewards^[1] for reinforcement learning^[2] rather than relying only on final success labels.

The authors analyzed approximately 100,000 real-robot execution records, including failures. They infer relative progress between nearby states instead of assigning absolute progress scores to every frame. Across 40 tasks in four LIBERO suites, average policy success improved by 18.16 percentage points against a sparse-reward baseline.

On real robots after 30 minutes of training, average success reached 60% on three precision-insertion tasks and 90% on three object-manipulation tasks. These are author-reported results, not a guarantee of safe online learning in factories.