fanqi-robo/diffusion_insert_gear_in_gripper_lr0p5x_s1000

diffusion 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	from scratch, except ImageNet-pretrained ResNet18 image encoders (one per camera)
Trainable	the whole model
Action chunk	horizon 64, 32 actions executed, 2 observation steps
Normalization	state and action min/max from the training set; images with ImageNet mean/std, resized to 576x1024 inside the encoders
Embodiment-specific pretraining	none (no robot pretraining at all)
Optimizer and schedule	Adam, diffusers cosine schedule with 500 warmup updates, reaching 0 at the last update
Learning rate	0p5x of the starting rate: optimizer_lr=5e-05
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, 13.2 GPU-hours
Parameters	293,174,382 trainable of 293,174,382
Peak VRAM	71.1 GB
Validation loss	0.0316369 at the final update; best 0.0143419 at update 6808
Training loss (last window)	0.00398349
Inference latency ms	255.1
W&B	https://wandb.ai/fanqi-robo-saferobotics/insert_gear_in_gripper_benchmark/runs/8qv7yul2

Validation loss: DDPM noise-prediction MSE of policy.forward in eval mode, mean of 4 noise draws per batch under a fixed seed; the encoders keep BatchNorm, so eval mode uses running statistics and the value is not the training-mode loss. 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.

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	25.96	30.90 +/- 2.95	26.03	36.17	0.85	0.43	3.6
1702	5.21	7.04 +/- 0.23	4.40	9.32	0.93	0.33	5.1
2553	3.45	4.80 +/- 0.34	2.86	6.57	0.93	0.33	4.1
3404	3.23	4.50 +/- 0.47	2.83	6.28	0.95	0.35	4.9
4000	2.65	3.57 +/- 0.42	2.33	4.85	0.95	0.29	4.8
4255	2.70	3.72 +/- 0.43	2.31	5.13	0.95	0.33	4.7
5106	2.59	3.63 +/- 0.34	2.16	5.05	0.96	0.40	4.6
5957	2.25	3.12 +/- 0.33	1.95	4.36	0.94	0.38	4.9
6808	2.23	3.34 +/- 0.49	1.72	4.90	0.92	0.46	4.5
7659	2.54	3.61 +/- 0.87	2.02	5.02	0.97	0.52	4.8
8000	2.21	3.14 +/- 0.38	1.77	4.38	0.95	0.57	4.8
8510	1.90	2.79 +/- 0.34	1.57	4.07	0.93	0.37	4.5
9361	1.99	2.82 +/- 0.46	1.63	4.00	0.95	0.59	4.2
10212	1.85	2.62 +/- 0.53	1.50	3.72	0.91	0.48	6.2
11063	1.91	2.82 +/- 0.69	1.51	4.08	0.90	0.56	4.7
11914	1.87	2.70 +/- 0.43	1.50	3.90	0.96	0.51	4.5
12000	1.77	2.62 +/- 0.54	1.41	3.84	0.93	0.51	4.5
12765	1.65	2.37 +/- 0.36	1.33	3.40	0.95	0.48	5.7
13616	1.77	2.64 +/- 0.65	1.35	3.81	0.95	0.62	4.4
14467	1.64	2.46 +/- 0.44	1.25	3.58	0.94	0.54	4.4
15318	1.52	2.24 +/- 0.40	1.18	3.20	0.96	0.52	4.7
16000	1.56	2.35 +/- 0.48	1.17	3.42	0.93	0.56	5.5
16169	1.59	2.39 +/- 0.55	1.20	3.45	0.95	0.52	4.4
17020	1.53	2.29 +/- 0.52	1.18	3.31	0.95	0.54	5.4
17871	1.48	2.23 +/- 0.48	1.11	3.24	0.94	0.57	4.7
18722	1.49	2.26 +/- 0.43	1.12	3.31	0.94	0.56	4.5
19573	1.48	2.26 +/- 0.49	1.10	3.32	0.94	0.56	4.3
20000	1.49	2.27 +/- 0.48	1.10	3.32	0.94	0.56	4.3