DiscoDemo generates diverse robot demos to improve imitation learning
Original title: DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning
Authors report downstream success rising from 40% to 62% in simulation and from 39% to 78% on a real robot.
DiscoDemo trains a privileged latent-conditioned reinforcement-learning generator to produce diverse successful robot demonstrations for later visuomotor imitation. A parallel reverse curriculum and success-gated diversity objective seek multiple solution modes; outputs are safety-filtered and exported as a static dataset. Across four manipulation tasks, the authors report broader successful-state coverage and higher downstream success for a pretrained flow-matching vision-language-action policy in simulation and on a real robot.
Unconfirmed: Coverage and success figures are the authors' reported results, not independently verified here.
- Averaged simulation success rose to 62% from 40% for the strongest baseline.
- Real-robot success rose to 78% from 39%, the authors report.
- The generator is not intended as a deployable controller.
- Authors say solution diversity remains complementary to corrective data collection.
From the original article
1 min read at huggingface.co
Behavior cloning for robotic manipulation is highly sensitive to the coverage of its training data, yet collecting diverse successful demonstrations remains costly. We study reinforcement learning (RL) not as a deployable controller, but as a data generator for downstream visuomotor imitation learning.
By Minho Park, Kinam Kim, Donghu Kim · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.