arkojit1/pi05_mj_stack_red_yellow_12.4k

π₀.₅ fine-tuned on arkojit1/mj_stack_red_yellow — best eval, step 12,400

lerobot/pi05_base fine-tuned with LeRobot (--policy.type=pi05) on arkojit1/mj_stack_red_yellow at revision d64e37c (the 2026-10-16 upload): 100 MuJoCo Franka block-stacking episodes, 8 language tasks (red-on-yellow and yellow-on-red, four phrasings each), two 224×224 cameras (front, wrist; the third π₀.₅ image slot is padded and masked, empty_cameras=1), 4-D state [x, y, z, gripper] (world-frame end-effector position + gripper opening), 4-D action [dx, dy, dz, gripper] (action.ee.delta_xyz + action.gripper.target).

Only data/chunk-000/file-000.parquet of that revision belongs to it. The other parquet files in the repo are left over from an earlier upload and were removed before training.

This is the run's lowest-eval checkpoint, step 12,400 (~41 epochs). Eval loss plateaued around 0.038–0.040 from ~16 epochs onward, and the run's final checkpoint is at arkojit1/pi05_mj_stack_red_yellow_20k (eval 0.0418). The next-best points were 0.0353 @8,200 and 0.0357 @7,400, so this one is within the eval noise of its neighbours rather than a sharp optimum.

Trainable	action expert only (train_expert_only=true; SigLIP + Gemma-2B frozen)
Hardware / batch	8 × MI300X, global batch 128 (16 per GPU), torch.compile (mode default)
Steps	12,400 of 20,000 (304 steps per epoch)
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	LeRobot default image transforms on train frames
chunk_size / n_action_steps	50 / 50
Eval split	last 10% of each task's episodes held out (38,932 training frames)
Eval loss	0.0347 (run best) at this step (flow-matching loss on held-out episodes)
Train loss	~0.016

Eval loss is a training-objective number on held-out simulator episodes, not a task success rate.

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("arkojit1/pi05_mj_stack_red_yellow_12.4k")