fanqi-robo/pi0_insert_gear_in_gripper_lr0p5x_s1000

pi0 on insert_gear_in_gripper (learning rate 0p5x, seed 1000)

This branch (main) holds the final checkpoint (update 20000). main is the final checkpoint and the benchmark's comparison point; best is the checkpoint with the lowest validation loss.

Task	insert_gear_in_gripper, bimanual YAM, 14-D joint state and action, three 720x1280 cameras
Training data	fanqi-robo/insert_gear_in_gripper @ acdc9ac8 (50 episodes, 27,228 frames)
Validation data	villekuosmanen/insert_gear_in_gripper_val @ 7c4d62f3 (5 held-out episodes, 2,245 frames)
Initialization	lerobot/pi0_base@26b99b94 (pretrained VLA: PaliGemma backbone and action expert)
Trainable	the whole model (freeze_vision_encoder=false, train_expert_only=false), bfloat16 with gradient checkpointing
Action chunk	50 actions predicted, 50 executed
Normalization	state and action mean/std from the training set; images rescaled in the model
Embodiment-specific pretraining	none known for bimanual YAM
Optimizer and schedule	AdamW, cosine decay with warmup; preset 1000/30000 rescaled by lerobot to warmup 666 / decay 20000
Learning rate	0p5x of the starting rate: optimizer_lr=1.25e-05, scheduler_decay_lr=1.25e-06
Updates	20000 optimizer updates at effective batch 32 (32 x 1, one GPU)
Seed	1000
Image augmentation	none
lerobot	0.5.1
GPU	1 x NVIDIA H100 80GB HBM3, 13.5 GPU-hours
Parameters	4,028,019,472 trainable of 4,028,019,472
Peak VRAM	44.4 GB
Validation loss	0.216177 at the final update; best 0.0659696 at update 851
Training loss (last window)	0.00913936
Inference latency ms	227.6
W&B	https://wandb.ai/fanqi-robo-saferobotics/insert_gear_in_gripper_benchmark/runs/k5n551wd

Validation loss: flow-matching MSE of policy.forward in eval mode, one noise draw per batch under a fixed seed; padded action steps count as targets (pi0 has no padding mask). It ranks checkpoints of this run; it is not comparable across policies.

benchmark_run.json holds the benchmark settings, metrics.jsonl the training and validation curves, run_summary.json the end-of-run numbers.

Licence

Fine-tuned from lerobot/pi0_base; use is subject to the Gemma Terms of Use of the base model.

Offline evaluation on the held-out episodes

Open-loop action-chunk error on villekuosmanen/insert_gear_in_gripper_val (423 queries, every 5th frame), in the dataset's joint units; hold (keeping the current pose) scores MAE@30 2.60.

Update	MAE@10	MAE@30	k=1	k=30	arm	grip rec	grip dt
851	1.31	2.17 +/- 0.38	0.83	3.36	0.95	0.46	4.5
1702	1.28	2.15 +/- 0.68	0.79	3.26	0.97	0.51	4.7
2553	1.15	1.90 +/- 0.46	0.77	2.94	0.95	0.60	4.2
3404	1.09	1.79 +/- 0.42	0.71	2.77	0.94	0.63	4.7
4000	1.08	1.81 +/- 0.46	0.72	2.85	0.97	0.68	3.8
4255	1.08	1.78 +/- 0.48	0.70	2.77	0.97	0.68	4.1
5106	1.02	1.67 +/- 0.42	0.67	2.60	0.99	0.62	3.1
5957	1.07	1.78 +/- 0.54	0.70	2.76	0.96	0.70	3.7
6808	1.10	1.84 +/- 0.61	0.71	2.84	0.96	0.59	3.2
7659	1.07	1.77 +/- 0.50	0.69	2.75	0.95	0.71	3.4
8000	1.04	1.74 +/- 0.46	0.68	2.71	0.95	0.67	3.1
8510	1.02	1.70 +/- 0.46	0.65	2.66	0.97	0.70	3.3
9361	1.03	1.74 +/- 0.51	0.66	2.74	0.96	0.68	4.3
10212	1.01	1.72 +/- 0.48	0.65	2.70	0.96	0.71	4.0
11063	1.01	1.72 +/- 0.49	0.64	2.70	0.96	0.70	3.8
11914	1.02	1.74 +/- 0.53	0.64	2.74	0.96	0.70	4.3
12000	1.00	1.73 +/- 0.54	0.63	2.74	0.97	0.70	3.7
12765	1.01	1.72 +/- 0.51	0.64	2.70	0.95	0.71	4.2
13616	1.02	1.72 +/- 0.48	0.66	2.71	0.96	0.70	3.5
14467	0.98	1.68 +/- 0.46	0.62	2.65	0.96	0.70	3.6
15318	1.00	1.70 +/- 0.49	0.63	2.68	0.96	0.68	3.9
16000	1.00	1.69 +/- 0.47	0.63	2.66	0.96	0.70	3.6
16169	1.00	1.70 +/- 0.50	0.63	2.68	0.96	0.70	3.9
17020	1.00	1.70 +/- 0.49	0.63	2.68	0.96	0.68	4.1
17871	1.00	1.70 +/- 0.50	0.63	2.68	0.96	0.68	4.2
18722	0.99	1.68 +/- 0.46	0.64	2.65	0.96	0.68	3.9
19573	0.99	1.68 +/- 0.47	0.62	2.66	0.96	0.70	3.9
20000	0.99	1.68 +/- 0.46	0.63	2.65	0.96	0.68	3.8