DiscoDemo Uses RL to Generate Diverse Robot Demos, Boosting Real-World Success from 39% to 78%
DAVIAN-Robotics · hf · 2026-10-27
KAIST and Holiday Robotics' DiscoDemo treats reinforcement learning as a data generator rather than a deployable controller to improve imitation learning:
- Idea: multiple successful solutions cover a broader success-trajectory space than repeatedly collecting one narrow solution; a privileged latent-conditioned RL generator with a parallel reverse curriculum and success-gated diversity objective discovers diverse solutions while preserving task success
- Pipeline: generated closed-loop behaviors are safety-filtered and exported as a static dataset for downstream visuomotor imitation learning
- Results: across four manipulation tasks, downstream success rose from 40% to 62% in simulation and from 39% to 78% on a real robot
More from Embodied
- 8.3M words transcribed across nearly 5,000 conversations on MIRA smart glasses — AnhPhuNguyen1 · 2026-10-27
- Surface Laptop Ultra review: impressive local AI but bandwidth limited — draginol · 2026-10-27
- Feather Robotics sells a $29,990 humanoid platform, betting on anti-overengineering — darian314 · 2026-10-27
- GPT-6-Astra embodied intelligence eval: near-100% on several of 12 benchmarks — chris_j_paxton · 2026-10-27
- John Carmack: no robot today can change batteries in a jack-o-lantern — ID_AA_Carmack · 2026-10-27
- Companion robot Moflin learns to comfort via local model and JEPA world model — Stefania_druga · 2026-10-27