fanqi-robo/pi0_insert_gear_in_gripper_lr2x_s1000

pi0 on insert_gear_in_gripper (learning rate 2x, 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	2x of the starting rate: optimizer_lr=5e-05, scheduler_decay_lr=5e-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, 13.7 GPU-hours
Parameters	4,028,019,472 trainable of 4,028,019,472
Peak VRAM	44.4 GB
Validation loss	0.249957 at the final update; best 0.0658975 at update 1702
Training loss (last window)	0.0070831
Inference latency ms	240.4
W&B	https://wandb.ai/fanqi-robo-saferobotics/insert_gear_in_gripper_benchmark/runs/vzfvh582

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	1.65	2.65 +/- 0.59	1.17	4.13	0.94	0.60	4.3
1702	1.52	2.26 +/- 0.32	1.16	3.28	0.92	0.56	4.9
2553	1.64	2.56 +/- 0.90	1.13	3.73	0.93	0.56	4.5
3404	1.28	1.96 +/- 0.33	0.98	3.00	0.91	0.60	5.1
4000	1.33	2.07 +/- 0.37	0.97	3.19	0.95	0.63	5.0
4255	1.37	2.26 +/- 0.63	0.93	3.46	0.96	0.79	4.7
5106	1.29	2.14 +/- 0.56	0.89	3.32	0.96	0.75	4.1
5957	1.34	2.10 +/- 0.52	0.95	3.19	0.92	0.75	4.7
6808	1.36	2.22 +/- 0.66	0.93	3.39	0.90	0.68	4.5
7659	1.31	2.18 +/- 0.70	0.86	3.36	0.97	0.73	3.3
8000	1.37	2.27 +/- 0.74	0.90	3.46	0.97	0.75	4.3
8510	1.24	2.08 +/- 0.51	0.84	3.26	0.94	0.78	4.3
9361	1.22	2.09 +/- 0.66	0.80	3.30	0.97	0.76	4.1
10212	1.25	2.13 +/- 0.76	0.80	3.31	0.98	0.75	3.5
11063	1.27	2.15 +/- 0.71	0.83	3.34	0.98	0.75	3.4
11914	1.27	2.16 +/- 0.77	0.83	3.35	0.95	0.75	3.5
12000	1.26	2.14 +/- 0.80	0.80	3.32	0.95	0.76	3.4
12765	1.24	2.11 +/- 0.78	0.78	3.28	0.97	0.76	3.2
13616	1.19	2.05 +/- 0.67	0.77	3.24	0.98	0.78	3.7
14467	1.20	2.06 +/- 0.71	0.76	3.24	0.97	0.78	3.9
15318	1.19	2.06 +/- 0.72	0.76	3.24	0.97	0.78	4.0
16000	1.19	2.06 +/- 0.72	0.75	3.25	0.98	0.76	3.6
16169	1.22	2.09 +/- 0.76	0.78	3.26	0.98	0.75	4.0
17020	1.18	2.04 +/- 0.71	0.75	3.21	0.98	0.73	3.7
17871	1.18	2.04 +/- 0.74	0.74	3.22	0.98	0.73	3.6
18722	1.18	2.04 +/- 0.71	0.75	3.22	0.98	0.73	3.7
19573	1.17	2.04 +/- 0.73	0.73	3.22	0.99	0.73	3.8
20000	1.17	2.03 +/- 0.71	0.73	3.21	0.98	0.75	3.8