Eval sets: MP tasks record only head cameras (front_stereo videos and their info/stats entries removed), BendPickAndPlaceOnSofaMP videos re-encoded to h264, drop BendPickAndPlaceOnSofaMP dr-level-hard-old and BendPickAndPlaceTeleop videos.pre-yaw0
Same fix as USC-PSI-Lab/SIMPLE-eval fb67cd9a: the policy instruction comes from environment_config.dr_state_dict.language, which differed from the training data.
- OpenFaucetTeleop: 'move forward to the faucet and open it' -> 'turn the faucet'
- XMovePickTeleop: 'bend to pick up the {}' -> 'move forward to pick up the {}'
Only meta/episodes.jsonl and meta/tasks.jsonl of each dr-level changed inside the zips.
18 days ago
eval-fixed BendPickAndPlaceOnSofaMP: move robot spawn 5-10 mm back from the target in 24 episodes with pre-grasp hand-target contact during the squat
BendPickAndPlaceOnSofaMP: grasp spawn offset (1 cm back, 1 cm left) in train and eval; eval dr-level-2 / dr-level-hard boxes moved toward the training placement; new dr-level-hard
dr_cfgs.language and dr_state_dict.language in dr-level-0/1/2 meta/episodes.jsonl still held the old 'bend the robot and pick up the {}, then place it on the table/container.' template, so the eval sent an instruction that differs from the episodes' tasks and the training data. Only those two fields change; all other zip entries are byte-identical.
eval-fixed: shifted-spawn eval levels (dr-level-0/1/2/hard) for 12 teleop tasks