NAJY_act_all11_hot_27D_120k_s2000

Same as the s1000 model with seed 2000. Offline it is the better seed on manipulation stages.

• Policy: ACT (lerobot 0.4.1 training, runs on 0.4.4), 52M params, chunk_size=30, n_action_steps=30, no temporal ensembling.
• Robot: Trossen Mobile AI (mobileai_robot), 3 cameras (cam_high, cam_left_wrist, cam_right_wrist, 640x480).
• Action: 16-D = 14 arm (left 6 + carriage, right 6 + carriage) + base x.vel, theta.vel.
• State: 27-D = 14 arm + 2 base slots trained as zeros + stage one-hot (K=11). Measured base velocity is never fed to the policy.
• Source checkpoint: /raid/kiro-ai/outputs/act/exp_all11_hot_s2000/checkpoints/120000 (DGX_1), trained with trossen-ai-simulation/scripts/train_multi.py (base_state: zero, task_onehot, balance, no augmentation, lr 3e-5, effective batch 32).

Running it on the robot

The robot emits a 14/16-D state, so this checkpoint needs the stage one-hot patch in Najongs/lerobot_trossen main (LEROBOT_TASK_ONEHOT):

LEROBOT_TASK_ONEHOT=<stage>/11 uv run lerobot-record \
  --policy.path=<this repo, downloaded> \
  --policy.n_action_steps=30 \
  --robot.type=mobileai_robot --robot.enable_base_motor_torque=true \
  ... --dataset.fps=21

• Stop the run if the log lacks LEROBOT_TASK_ONEHOT=... installed and stage i/11 active.
• Do not pass --policy.type or --policy.temporal_ensemble_coeff.
• One stage per run. Stage numbers:

stage	dataset	LEROBOT_TASK_ONEHOT
task01	move_to_tube_rack	1/11
task02	pickup_tubes	2/11
task03	turn_to_face_beaker	3/11
task04	pour_liquid_from_tubes_to_beaker	4/11
task05	tube_disposal	5/11
task06	pickup_beaker_and_move_to_refrigerator	6/11
task07	open_refrigerator	7/11
task08	takeout_and_put_beaker	8/11
task09	close_refrigerator	9/11
task10	move_to_beaker_shelf	10/11
task11	place_beaker_on_the_shelf	11/11

Offline results (not robot results)

Locomotion stages, base per-step MAE (lower is better), exec=5, teacher-forced, 24 episodes: task01 0.0251 (s1000) / 0.0307 (s2000) · task03 0.0347 / 0.0379 · task06 0.0197 / 0.0212 · task10 0.0507 / 0.0546 — tie with the locomotion-only model (mean −0.0019). Manipulation arm MAE (rad): task02 0.0333 / 0.0285 · task07 0.0467 / 0.0351 · task08 0.0349 / 0.0186 — 3.0–10.2× the seed spread better than per-stage ACT.

All numbers are offline replays of held-out (time-ordered) episodes; none of this has run on the robot yet. exec (re-query period) is not decided: teacher-forced replay favours exec=5, a replay closed-loop proxy favours exec=30 — use 30 until a robot A/B. Details: trossen-ai-simulation docs/mobile_base_investigation.md §82–86, docs/real_robot_eval.md.