Ο€β‚€.β‚… fine-tuned on arkojit1/mj_stack_red_yellow_3cam β€” batch 64, step 20k of 20k

lerobot/pi05_base (revision a538eb27) fine-tuned with LeRobot 0.7.0 on arkojit1/mj_stack_red_yellow_3cam (revision a38bf81): 200 MuJoCo Franka (panda) block-stacking episodes at 50 Hz, 8 language tasks (four phrasings each of "red on yellow" and "yellow on red", 25 episodes per task).

  • observation.state (4-D): end-effector x, y, z + gripper
  • action (4-D): [dx, dy, dz, gripper], the cartesian_pos scheme: the dataset's action.ee.delta_xyz followed by action.gripper.target. The dataset's orientation and joint-delta action columns were not used, so the policy never commands rotation.

The policy outputs a 50-step chunk (1.0 s) of these; translation deltas are as stored in the dataset and must be applied the same way, in this dimension order, at deployment.

Cameras. The policy's image inputs are the dataset's own keys, in this order: observation.image.front, observation.image.wrist, observation.image.side. Ο€β‚€.β‚… fills its image slots in that order, so front takes base_0_rgb, wrist left_wrist_0_rgb, and side right_wrist_0_rgb (empty_cameras=0). Pass all three under their dataset names.

This is the final (and best) checkpoint, step 20,000 (15.3 epochs) of a 20,000-step run; it is the last step and has the run's lowest eval loss. Eval was still descending slowly at the end of the schedule, from ~0.018 over steps 13,000–19,000.

Trainable action expert only (train_expert_only=true; SigLIP + Gemma-2B frozen)
Hardware / batch 8 Γ— MI300X (8 per GPU), global batch 64
Steps 20,000 planned (1310.9 steps per epoch = 15.3 epochs)
Optimiser AdamW, LR 2.5e-5 peak, betas (0.9, 0.95), wd 0.01, grad clip 1.0
Schedule cosine with warmup, pinned to 20,000 steps (warmup 666), floor 2.5e-6; bf16
Normalisation quantile (q01–q99) for state and action
Augmentation none
chunk_size / n_action_steps 50 / 50
Eval split last 3 episodes of each of the 8 tasks (24 of 200) held out; trained on the other 176
Eval loss 0.0170 at this step (flow-matching loss on held-out episodes)
Train loss 0.011

Eval loss is a training-objective number on 24 held-out episodes, not a task success rate; point-to-point scatter is about Β±0.001 late in the run.

LeRobot version. Trained and saved with LeRobot 0.7.0 (stock transformers 5.22). Its Ο€β‚€.β‚… prompt carries the state as only the real (4) discretized dimensions, not padded to 32 as in LeRobot 0.4.4, so load and run it with LeRobot 0.7.

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("arkojit1/pi05_mjstack3cam_bs64_20k_20k")
Downloads last month
29
Safetensors
Model size
4B params
Tensor type
F32
Β·
BF16
Β·
Video Preview
loading

Model tree for arkojit1/pi05_mjstack3cam_bs64_20k_20k

Finetuned
(913)
this model

Dataset used to train arkojit1/pi05_mjstack3cam_bs64_20k_20k