The current leaders.
Success rate across 20 tasks.
Every policy, every metric.
Click a heading to sort.
Ranked by success rate, ties broken by mean score. Click a column heading to re-rank; hover a bar for its confidence interval.
Same conditions for everyone.
Every policy runs under the same conditions, so differences in the table come from the policy rather than the setup.
Official checkpoints
Each policy uses its official DROID-finetuned checkpoint, with no adaptation to our environments.
Shared setup
Absolute joint-position control, and the same simulation parameters, robot configuration and camera setup for every policy.
Six rollouts per task
Each of the 20 tasks is attempted six times, with 1200 simulation steps per rollout.
Success and progress
Evaluators return a binary success verdict and a continuous progress score, which the table reports as success rate and mean score.