UC Berkeley and 17 Labs Open-Source Robot Failure Dataset OopsieData

UC Berkeley, together with scholars including Sergey Levine and Peter Stone, and 17 laboratories, launched the OopsieData open-source dataset initiative to collect real-world suboptimal operations and failure cases of robots. The data is now fully open to the public, aiming to enhance reinforcement learning and agent capabilities by learning from failures, and calls for more labs to join and share failure data.

Confirmed

• The project is initiated by UC Berkeley, Sergey Levine, Peter Stone, and other scholars, with 17 labs contributing data.
• The dataset includes not only successful operations but also deliberately collected suboptimal solutions and short grasp-failure videos, which are the richest yet underutilized resource in robot learning.
• This project follows the team's previous work on OopsieVerse, a damage-aware simulator, marking another advance in robot safety and intelligence research.

Why it matters

• Traditional RL and agent training often rely on imitating successful cases; OopsieData fills this gap. According to @PeterStoneTX, the project aims to improve intelligence by enabling models to predict and avoid actions that may lead to failure.
• The team has opened collaboration, urging more labs to share failure data to address long-standing challenges in real-world robot operation training.

2026-07-31 ~ 2026-08-01 · 5 related posts

• OopsieData: Turning Discarded Robot Grasping Failure Clips into Training Data — chris_j_paxton · 2026-07-31
• [source] Oopsie-Data Project Opens for Collaboration: Learning from Failures in RL — PeterStone_TX · 2026-07-31
• OopsieData: The First Large Dataset of Real-World Robot Failures — RobobertoMM · 2026-07-31
• 17 Labs Launch OopsieData: Open-Sourcing Real-World Robot Manipulation Data — SongShuran · 2026-08-01
• Berkeley Launches OopsieData: An Open Dataset for Robot Failure Clips — berkeley_ai · 2026-08-01