Oopsie Data: Multi-Lab Effort Builds First Large-Scale Dataset of Real Robot Failures
tw_killian · x · 2026-10-23
Oopsie Data is a multi-lab effort to build the first large-scale dataset of real-world robot manipulation failures. Today's robotics datasets contain only successes, yet a policy that has only seen things go right never learns what a bad grasp looks like.
- Failure scenarios are produced constantly during policy rollouts in every lab but thrown away; real failures are the missing ingredient for RL, reward modeling, failure prediction, and world modeling, and synthetic noise injection is no substitute.
- The ask: next time you roll out a policy on a real robot (evaluation, play data, online RL), keep and submit both failures and successes.
- Contributing labs get early access and can qualify for co-authorship on the public release.
- Initial examples show a diffusion policy on an Aloha robot: a slight gripper offset turns a ball grasp into failure—the fine-grained differences the dataset aims to capture.
More from Embodied
- KAIST's FRED lifts robot policy success from 19.4% to 50% via self-distilled future video representations — kaist-ai · 2026-10-23
- Figure prints 81 minutes of robot data per second, spent ~$8.1M on data last week — joecole · 2026-10-23
- Just three physical AI agents are actually paid for at scale, argues investor — jokrvivek · 2026-10-23
- UCLA roommates-turned-founders: the story behind sidewalk delivery robot startup Coco Robotics — zhoubolei · 2026-10-23
- Figure 04 hits full design lock; Brett Adcock calls it the iPhone 1 of humanoids — chris_j_paxton · 2026-10-23
- Position paper maps recursive self-improvement in physical AI, finds no study isolates recursive gains — NationalUniversityofSingapore · 2026-10-23