The 5 MP tasks (g1_wholebody) replayed 4/100 (TabletopHandoverMP) after the Dex3 hand PD change (kp 40-80 -> 5/2.5, index/middle torque 0.7). Changes:
- Hand actions boosted: a' = a + (kp_old/kp_new - 1) * deadband(a - q_next), clipped to joint limits; only stall errors pushing away from the open pose (not releases). BendPickAndPlaceOnSofaMP: no boost on thumb_0 or the left hand. Settings per task in handboost.json.
- PickAndBendPlaceMP: 14-frame right-hand release appended to 10 episodes whose recording ended with the hand closed.
- PickAndBendPlaceMP, BendPickAndPlaceOnSofaMP: 20 frames holding the final row appended to every episode (released box lands inside the episode).
- Episodes failing the MuJoCo replay removed, rest renumbered: TabletopHandoverMP 76/100, BendPickMP 91/100, TabletopGraspMP 98/100, PickAndBendPlaceMP 49/57, BendPickAndPlaceOnSofaMP 75/99. stats.json / stats_psi0.json are those of the full datasets.
PickAndBendPlaceMP 49/49 and BendPickAndPlaceOnSofaMP 75/75 r…
data-fixed: fix 1-frame lag in re-rendered G1WholebodyBendPickMP-v0 videos
Fix off-by-one in the previous re-render commit: the replay video starts with the reset frame, but dataset frame k shows the state after action k, so the videos were one frame behind the states/actions. Now uses render frames [1, 1+length): frame-difference correlation with the original videos peaks at shift 0 for 100/100 episodes (was +1). Only videos change; codec, resolution, fps, frame counts and timestamps unchanged.
data-fixed: re-render G1WholebodyBendPickMP-v0 videos with per-episode table materials
Egocentric videos of all 100 episodes re-rendered with the current renderer so the table/ground materials match each episode's environment_config (as in eval). 100/100 replays succeeded. States, actions, meta and all non-video files are byte-identical; videos keep the same codec (h264 High, yuv420p), 640x360, 50 fps, frame counts and timestamps.
data-fixed: MP waist state fix, recomputed stale stats
MP tasks (BendPickAndPlaceOnSofaMP, BendPickMP, TabletopGraspMP, TabletopHandoverMP): states[28:31] and
observation.prev_torso_rpy now follow the AMO convention, last commanded waist roll/pitch/yaw =
action[t-1, 28:31] (0 at frame 0), as the psi0 eval agent (simple/baselines/psi0.py) feeds them.
Actions, videos and states[31] unchanged. PickAndBendPlaceMP already followed it (not touched).
meta/stats.json and stats_psi0.json recomputed from the data where stale (those 4 MP tasks and BendHandover,
BendPickAndPlace, BendPickTeleop, LocomotionPickBetweenTables, XMovePick); meta/episodes_stats.jsonl
rows whose count did not match the episode length recomputed (433 rows in the 5 teleop tasks).
20 days ago
data-fixed: G1WholebodyBendHandoverTeleop-v0 rebuilt with the full teleop tails (90 eps +2357 rows, 40775 frames; video pads the last frame; stats unchanged)
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