Policy Leaderboard
26Tasks12Embodiments4Channels
Evaluation settings vary by method. Protocol & metrics ↓
Metric
Policies
All policies
Evaluation Protocol & Metrics
- SR Stable Success Rate
- Episodes reaching stable success (%), averaged equally across tasks and four channels.
- LSCR Latched Stage Completion Rate
- Dependency-valid task stages ever reached (%), averaged equally across tasks and four channels.
- Eff. Success-Conditioned Efficiency
- Mean expert completion steps / mean policy steps to stable success, pooled across successful episodes. Averaged over tasks with success.
Bench2Dex evaluations use 50 episodes per task and channel. GPT-6-Astra uses ICL and 3 episodes per task on Full only; 1× / 10× denote execution-step budgets. Full SR is not a four-channel mean. Phoenix protocol ↗ · GPT-6-Astra protocol ↗
