Deployment often exposes long-tail failures that fixed demonstration datasets cannot fix. This paper's core insight is that deployed robots themselves are a distributed data source: by sharing execution traces and using failure contexts to retrieve relevant successful segments, a common policy can be fine-tuned online at fleet scale without additional expert corrections.
Key Findings
- Failure-guided Experience Retrieval Fine-Tuning (FERT): when a robot fails, the system retrieves matching successful trajectory segments from the fleet and uses the actions from those successes as supervision for policy updates—so it learns corrective behavior without collecting new expert demonstrations.
- Fault-tolerant Asynchronous Training System (FATS): coordinates experience collection, optimization, and policy distribution across many robots while tolerating local interruptions and heterogeneous availability, enabling continuous learning at scale.
- Strong empirical gains and scaling: across five real-world manipulation tasks the paper reports 93.3–99.2% success on held-out robots (120 trials per task). With fixed collection time per robot, larger fleets yield more experience and better final performance—for multi-stage phone packaging, growing the fleet to 96 robots raised success from 79.2% (initial) to 98.3%.
Who it's for and tradeoffs
Great fit if you operate or design multi-robot deployments and want a practical mechanism to improve policy reliability using only execution traces—especially for manipulation-heavy workflows where collecting teleoperated corrective demonstrations is costly. Look elsewhere if you need per-robot personalized policies, tight formal safety guarantees for novel failures, or if your deployment cannot share experience data across agents for privacy/regulatory reasons.
Brief method overview
The approach combines retrieval-based supervision (match failure contexts to successful segments), supervised fine-tuning on retrieved action sequences, and an asynchronous cloud-backed pipeline that gathers trajectories, runs optimization, and rolls out updated policies back to robots. The loop emphasizes fault tolerance and fleet-wide aggregation rather than per-episode human-in-the-loop corrections.