DexArena separates failures in execution, generalization, and composition while preserving matching task definitions.
Run each of the 30 canonical stage environments with its registered robot, horizon, action layout, and physical success predicate.
Evaluate every stage on three native scene surfaces using identical object placement and the fixed 256 × 256 policy camera rig.
Execute two to five dependent stages in their parent long-horizon environment and report complete-sequence success.
Fraction of evaluation episodes whose stage predicate is satisfied for its required hold duration.
Mean stage success over robocasa_l3_s2, robocasa_l3_s1, and robocasa_l3_s6 with matched seeds.
Difference between the reference collection workspace and native-scene performance.
Fraction of complete long-horizon episodes reaching the final task success state.
All six baselines use image observations and robot proprioception, with matched task horizons, randomization and embodiments. Stage and long-horizon success are measured over 50 episodes per task; generalization uses 50 episodes in each of three held-out kitchens without further adaptation.
The manuscript specifies stage seeds 50000–50049 for ACT and DP. GR00T uses 10001–10050, reused for each generalization scene and the long-horizon protocol. Record the seed manifest alongside each result.
Generalization reports mean ± standard deviation across three scenes. The average row aggregates stage performance within each scene before reporting variation across scenes.
PPO, RLPD, ACT, DP and π0.5 switch between learned stage policies when the stage succeeds. GR00T uses one task model trained on pooled stage data, with a new instruction at each transition.
GR00T records the highest average success in the training scene. ACT and RLPD follow, but no method consistently solves every stage.
50 evaluation episodes per stage. Scores are averaged over 30 canonical stages.
π0.5 leads across three held-out kitchens. Its average is 2.9 percentage points below the training-scene result; the same policy is evaluated without adaptation.
50 episodes per stage per scene. ± denotes standard deviation across three scene averages, not a confidence interval.
π0.5 and ACT report non-zero average success on complete sequences. Successful individual stages do not yet translate into reliable end-to-end behavior.
50 episodes per complete task. Success requires reaching the final task condition without resets between stages.
