DSP turns 15 minutes of human play into robot dexterity, no demos or sim2real needed
anand_bhattad · x · 2026-10-17
Amazon Frontier AI & Robotics and University of Maryland (authors include Jitendra Malik and Ruoshi Liu) introduce Dexterity from Supervised Play (DSP): 15 minutes of goal-free glove teleoperation "play" is enough to learn both a world model and a dexterous manipulation policy, with no task demonstrations and no sim2real.
The five-stage pipeline:
- A diffusion play policy (diffusion transformer + behavior cloning) learns 22-joint motion chunks from the play data;
- The policy runs autonomously on the robot for unlimited data collection (scaling study across 0.54–5.11 h);
- A Diffusion Forcing world model predicts future cube poses from recent observations and candidate action chunks;
- At each step DSP samples 1,024 play-policy proposals, executes the one closest to the goal, and replans;
- Optionally, the model-based planning is distilled into a 21.8 ms goal-conditioned policy retaining over 90% of MPC success at 0.8 rad.
Paper and code are public.
More from Embodied
- Ocean robotics firm Ulysses runs multi-region missions, hires more robot deployers — espricewright · 2026-10-17
- European researchers build autonomous robots to patrol aging sewer systems — Polymarket · 2026-10-17
- Scoble heads to SF humanoid robotics conference, says absent Optimus is the elephant in the room — Scobleizer · 2026-10-17
- TANDEM mixes autonomous TAMP with on-demand teleoperation for 2.9× more demos — tomssilver · 2026-10-17
- DATAFARM gains a human-in-the-loop to handle what TAMP alone cannot — tomssilver · 2026-10-17
- Ex-Nvidia robotics lead: robots are in their "Lego era", no ChatGPT moment yet — TechCrunch AI · 2026-10-17