angelic123/robotwin_evaluation

RoboTwin evaluation: FastWAM stock vs. PhyPO

Raw evaluation outputs (logs, per-phase results, episode videos) of FastWAM (Wan2.1 1.3B, checkpoint step 93932), stock vs. PhyPO-trained, for 18 of the 50 RoboTwin dual-arm tasks: those where PhyPO's mean success rate over the clean and randomized settings is higher than stock's. Both models' outputs are included for each task. Settings:

• Clean (demo_clean): plain background, uncluttered table at a fixed height, constant lighting.
• Randomized (demo_randomized): random background texture (plain in 2% of episodes), distractor clutter on the table, table height perturbed by up to 3 cm, randomized lighting (extreme in 2% of episodes).

Each task x setting ("phase") is 100 episodes; instructions come from unseen templates; both models use the same seed sequence.

Layout

• FastWAM_runs/20261006_012857_stock_e100/, FastWAM_runs/20261006_012857_phypo_e100/: the main runs.
  • <task>/_result_clean.txt, <task>/_result_random.txt: final success rate of the phase.
  • <task>/episode<i>_randomized-<bool>_success-<bool>.mp4: one video per episode.
  • eval_<task>_<timestamp>.log: run logs (including retries after simulator render hangs).

Tasks: click_alarmclock, handover_block, open_laptop, open_microwave, pick_diverse_bottles, pick_dual_bottles, place_bread_basket, place_burger_fries, place_cans_plasticbox, place_dual_shoes, place_mouse_pad, place_object_basket, place_phone_stand, put_bottles_dustbin, stack_blocks_three, stack_bowls_three, stack_bowls_two, turn_switch. A few task folders also hold videos from earlier aborted attempts of the same phase.