Event
DiscoDemo: Generating Diverse Demonstration Data with Reinforcement Learning
1 reports · 1 sources
Minho Park and co-authors propose DiscoDemo, which uses a latent-conditioned reinforcement learning generator with privileged information, combined with parallel reverse curricula and a success-gated diversity objective, to produce multiple successful closed-loop behaviors that are then exported as a static dataset after safety filtering, for use in downstream visuomotor imitation learning. On four manipulation tasks, the authors report that downstream average success rates improve from 40% for the strongest baseline to 62% in simulation, and from 39% to 78% on real robots. The authors state that diversity of successful solutions can improve imitation learning of pretrained flow-matching VLAs and is complementary to corrective data collection; these effects are results reported by the authors.
Report timeline
- DiscoDemo: Using Reinforcement Learning to Generate Diverse Demonstration Data to Improve Imitation Learning
DiscoDemo uses a latent-conditioned reinforcement learning generator with privileged information, together with a parallel reverse curriculum and a success-gated diversity objective, to produce multiple successful closed-loop behaviors; after safety filtering, these are exported as a static dataset for downstream visuomotor imitation learning. On four manipulation tasks, the average downstream success rate in simulation rose from 40% for the strongest baseline to 62%, and on real robots from 39% to 78%. The authors believe that the diversity of successful solutions can improve imitation learning for pretrained flow-matching VLAs and complements corrective data collection.
arXiv cs.RO ·