SO-101 BENCH

Live leaderboard

916 episodes. Five task families. Three object splits.

Loading latest published results…
SO-101 Bench primary evaluation. Benchmark score with 95% confidence intervals. Other columns show raw subset success rates.
#PolicyEpisodesScore / 95% CISeenNew instanceNew classROBUSTPLANMEM
Loading paper results…

Score = mean success over 15 task-family × object-split cells, with a 95% within-cell bootstrap CI. Other columns show raw success rates.

Download results CSV ↓
Scoring and subsets

Benchmark score is the equal-weight mean over 15 task-family × object-split cells, with a 95% within-cell bootstrap confidence interval. All other columns show raw success rates.

ROBUST · 178 held-out appearance episodes. PLAN · 48 rotation-to-fit episodes. MEM · 51 occlusion episodes. Subsets overlap.

Published results refresh every 60 seconds. The paper figure retains the fixed paper snapshot.

Result history