NEEDLEwork stitches suboptimal robot demos into better training data
SongShuran · x · 2026-10-08
- Researchers introduce NEEDLEwork: rewriting robot demonstration data offline by inserting short, verified action bridges between recorded observations — no simulator, new robot time, or privileged state needed
- Core idea: imperfect demos contain useful behavior alongside repeated grasps, detours, and failures, and the best behavior often exists in no single trajectory; stitching lets a policy skip detours, borrow better continuations, and recover from failed attempts
- Three stitch types: S1 within-trajectory shortcut, S2 cross-trajectory transplant of complementary demos, S3 failure funnel connecting a failed state to a successful continuation
- Evaluated on real robots over 300 trials with success rates; paper and code are open
More from Embodied
- RLHND turns video diffusion models into physically grounded hand trackers for robot learning — RLWRLD · 2026-10-08
- Stanford's MobileVISTA fixes mobile manipulation's pose brittleness with generative data augmentation — jiajunwu_cs · 2026-10-08
- Researchers compile a larval fish brain into a simulated circuit running 26x faster than spiking models — mtizard · 2026-10-08
- US Treasury issues first outbound investment fine: $200k over $92k Chinese robotics AI deal — pstAsiatech · 2026-10-08
- Palmer Luckey: I Believe in Humanoid Robots, But Won't Build Them Myself — PeterDiamandis · 2026-10-08
- Adding Motion Sense to VLAs With Frozen SAM 2 Lets Robots Catch Moving Bottles — nikhilaravi · 2026-10-08