OopsieData: multi-lab effort builds first large-scale dataset of real robot failures
micoolcho · x · 2026-10-23
The robotics community is launching Oopsie Data, a multi-lab effort to build the first large-scale dataset of real-world robot manipulation failures. Today's robotics datasets contain only successes — but a policy that has only seen things go right never learns what a bad grasp looks like.
- Real failures and suboptimal behavior are the missing ingredient for RL, reward modeling, failure prediction, and world modeling; synthetic noise injection is no substitute
- Labs contributing rollouts (failures and successes alike) get early access and can qualify for co-authorship on the public release; a quickstart is provided
- The initial dataset includes paired success/failure episodes of an Aloha robot grasping a ball with a diffusion policy, showing how slight gripper offset causes failure
- A RoboPapers podcast episode discussing the project is dropping soon
More from Embodied
- Unitree G1 Drives a Go-Kart as Ex-Tesla Optimus Member Recalls Secretive Early Days — ATTlKA · 2026-10-23
- Chinese robotics companies' overseas sales heads often don't speak English, observer says — broodsugar · 2026-10-23
- Asimov ships new grippers and an extra arm DOF with third-party support — freelerobot · 2026-10-23
- NAVER AI Lab releases LaLa-World, scaling latent world models with natural-language actions — naver-ai · 2026-10-23
- SkeLAM grounds latent actions in robot skeletons for controllable world models — Qi Lv · 2026-10-23
- Ken Goldberg: humanoid lab success rates stalled at 70-80%, 'not acceptable for real industry' — AIFlow_ML · 2026-10-23