A robotics policy trained on 4 hours of diverse data beats in-domain data
DominiqueCAPaul · x · 2026-07-27
The author reports an unexpected robotics finding: a policy trained on 4 hours of data gathered across 5 tables and 2 workstations outperformed, on the lightbox evaluation, a policy trained on the same amount of data collected inside the lightbox itself.
The main takeaway is that diverse out-of-domain data beat in-domain data on an in-domain benchmark. The author says they expected hyperparameters to matter more, but data collection strategy turned out to be the bigger factor.
More from Embodied
- Qualcomm robot collapses mid-presentation in an awkward demo fail — iamfakhrealam · 2026-07-27
- Apollo Go starts its first fully driverless trial in Hong Kong — Baidu_Inc · 2026-07-27
- WorldDreamerV4 targets shared world models for robot swarms and tops RoboCasa, WorldScore — 机器之心 · 2026-07-27
- World Action Models survey says robotics is moving from reacting to predicting consequences — rohanpaul_ai · 2026-07-27
- AheadForm unveils an ultra-lifelike humanoid robot at WAIC 2026 in Shanghai — Olivier__OG · 2026-07-27
- Stanford AI Lab cites NVIDIA’s Alpamayo platform as a model for open Physical AI — StanfordAILab · 2026-07-27