Event
96 robots share experience, raising operation success rate to 98.3%
2 reports · 2 sources
Chao Yu et al. propose the "Collective Experience Learning" (CEL) paradigm: robots in deployment aggregate their own execution experience to update a shared policy, and the updated policy is then returned to the robot fleet for use. It is implemented as Failure-guided Experience Retrieval Fine-Tuning (FERT), which uses failure context to retrieve relevant segments from successful trajectories and prioritizes supervised updates using the actions recorded in these segments; a supporting fault-tolerant asynchronous training system (FATS) coordinates robot execution, cloud inference, and policy optimization. On five real manipulation tasks, the resulting policies achieved success rates of 93.3%–99.2%, with each task evaluated 120 times on robots that did not participate in training data collection. In a multi-stage phone packaging task, learning with 96 robots raised the success rate from 79.2% to 98.3%. The method presupposes initialization with expert supervision, as well as binary success/failure labels for each episode and environment resets provided by an operator; under these conditions, no additional expert demonstrations or corrective actions need to be collected. Given a fixed data collection time per robot, a larger fleet provides more experience, and FERT generally achieved higher success rates at the scales evaluated.
Developments
- Learning from Fleet Experience: Collective Experience Learning Across 96 Robots · 1 reports
- Learning from Fleet Experience: Improving Robot Manipulation Policies with Shared Fleet Experience · 1 reports
Report timeline
- Learning from Fleet Experience: Collective Experience Learning Across 96 Robots
The paper proposes the Collective Experience Learning (CEL) paradigm, in which deployed robots pool their own execution experience to update a shared policy that is then returned to the fleet for use. It is implemented through failure-guided experience retrieval fine-tuning (FERT), which uses failure contexts to retrieve relevant segments from successful trajectories and prioritizes their recorded actions for supervised updates, without additional expert demonstrations or corrective actions; a supporting fault-tolerant asynchronous training system (FATS) coordinates robot execution, cloud inference, and policy optimization.
arXiv cs.RO ·
- Learning from Fleet Experience: Improving Robot Manipulation Policies with Shared Fleet Experience
The paper proposes the collective experience learning (CEL) paradigm, in which deployed robots share execution experience to improve a common policy without collecting additional expert demonstrations or corrective actions. It uses failure-guided experience retrieval fine-tuning (FERT) to retrieve relevant trajectory segments from the fleet's shared successful experience as supervision, and a fault-tolerant asynchronous training system (FATS) to coordinate experience collection, policy optimization, and distribution.
HuggingFace Daily Papers ·