Robots don't fail in the lab. They fail in the field.
Smiling Buddha is an open, community-run initiative building the real-world data layer for physical AI: edge cases from real deployments, native to @LeRobotHF on @huggingface, free for the community.
smilingbuddha.co
CoRL week made the measurement problem obvious: LIBERO can be shortcut-solved, and as @junfanzhu98 notes, more data won't help if time and labels are misaligned.
Scoring a VLA? Add a real-deployment failure split.
teapawt.net/junfanzhu98/status/212…#PhysicalAI#CoRL2026
CoRL 2026 Day 1🤖 @junfanzhu98's Hot🔥 Takes
10 takeaways: shortcut-solvable benchmarks, VLAs that don't listen, video as policy, sim-to-real, time-aware memory, and self-improving agents.
#CoRL2026@corl_conf
Free for the community.
@LeRobotHF and @huggingface gave Physical AI a shared format and hub. We stay complementary: LeRobot v2.1 native corpora + V0–V3 language.
Not rivals. Same stack, different layer.
RSEF, our Real-to-Sim Enrichment Framework v1.0, is live.
Turn a narrow real capture into scored sim variants.
V0 synthetic base → V1 single real match → V2 multi-site → V3 novel discovery.
Not more sim. Enrichment with a score.
Lab demos look clean. Customer floors do not.
Physical AI still trains too often on lab and synthetic data, then meets lighting shifts, layout drift, and recovery mess in production.
We catalogue those edge cases from live deployments.