JackLiu0406/DA3-VLA-RoboPro-Results

DA3-VLA — RoboPRO Evaluation Results

Full evaluation results on the RoboPRO benchmark (RoboTwin / SAPIEN sim, 79 tasks across kitchenl / kitchens / office / study), organized by model. Each model folder contains the scored metrics, per-task CSVs, rollout videos, and the model checkpoint.

Protocol: stock TOPP control; clean = 20 seeds/task, clutter = d6–d15 × 2 seeds/task. Clean and clutter are reported separately (never averaged).

Headline numbers

Model	Clean SR / HSR	Clutter SR / HSR
pi05_da3_v2final_ckpt39999 — pi0.5 + DA3 (L6-Action XAttn), 30/50 exec	81.3 / 62.8	72.8 / 25.1
pi05_base_50-50 — vanilla pi0.5 (clean 50/50, clutter 30/50)	70.1 / 54.9	62.2 / 22.9
da3-xvla-45k — giant X-VLA + DA3 (L6-Action XAttn), ckpt 45k	55.6 / 46.7	45.3 / 11.2*

*da3-xvla-45k clutter is partial (n=1187, ~75%). The two pi05 models are complete and matched. DA3 beats vanilla pi0.5 on both configs: clean +11.2/+7.9, clutter +10.6/+2.2.

Layout (per model)

<model>/
  SUMMARY.txt              # TOTAL + per-scene + per-task, for CLEAN and CLUTTER sections
  per_task_clean.csv       # scene,task,SR,HSR,n
  per_task_clutter.csv     # scene,task,SR,HSR,n
  <scene>__<task>/         # rollout videos: episodeN.mp4 (clean), d<N>_episodeN.mp4 (clutter)
  checkpoint/              # the model weights used for this eval

Models:

• pi05_da3_v2final_ckpt39999 — pi0.5 + DA3 spatial addon, v2-final checkpoint (openpi/JAX).
• pi05_base_50-50 — base pi0.5 (jax 30k), the vanilla baseline.
• da3-xvla-45k — v1-nested-giant-perc-ckpt45000, giant X-VLA with DA3 (policy dxvla).

Notes

• SR = success rate; HSR = hard success (success with zero collisions).
• Sim: SAPIEN 3.0.0b1 on GB10 (aarch64). Long-horizon kitchens chains and dense clutter are the hardest cells; clutter HSR is collision-bottlenecked for all models.