fanqi-robo/diffusion_insert_gear_in_gripper_lr1x_s1000

diffusion on insert_gear_in_gripper (learning rate 1x, 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	1x of the starting rate: optimizer_lr=0.0001
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, 12.3 GPU-hours
Parameters	293,174,382 trainable of 293,174,382
Peak VRAM	71.1 GB
Validation loss	0.0339572 at the final update; best 0.0149422 at update 6808
Training loss (last window)	0.0038125
Inference latency ms	272.8
W&B	https://wandb.ai/fanqi-robo-saferobotics/insert_gear_in_gripper_benchmark/runs/r4urrden

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	21.22	25.09 +/- 4.07	20.14	29.41	0.87	0.41	5.6
1702	5.54	7.13 +/- 0.71	4.81	9.21	0.93	0.33	5.9
2553	3.27	4.53 +/- 0.45	2.73	6.21	0.96	0.38	3.8
3404	4.55	5.96 +/- 0.55	3.86	7.80	0.89	0.25	3.6
4000	3.04	4.06 +/- 0.64	2.64	5.35	0.98	0.32	4.3
4255	2.73	3.74 +/- 0.46	2.32	5.16	0.93	0.40	3.8
5106	2.51	3.40 +/- 0.27	2.08	4.61	0.98	0.44	5.2
5957	2.30	3.08 +/- 0.23	2.05	4.15	0.97	0.46	4.9
6808	2.41	3.63 +/- 0.43	1.93	5.34	0.96	0.56	3.7
7659	3.07	4.48 +/- 1.37	2.41	6.28	0.98	0.60	4.4
8000	2.11	2.82 +/- 0.39	1.76	3.75	0.99	0.54	4.3
8510	2.04	2.77 +/- 0.21	1.70	3.80	0.96	0.46	5.7
9361	2.09	2.90 +/- 0.48	1.72	4.01	0.95	0.56	3.8
10212	1.95	2.64 +/- 0.25	1.62	3.62	0.96	0.48	4.0
11063	1.99	2.82 +/- 0.73	1.59	3.98	0.96	0.65	3.9
11914	1.88	2.54 +/- 0.31	1.55	3.46	0.96	0.57	4.9
12000	1.76	2.41 +/- 0.26	1.43	3.28	0.98	0.56	4.7
12765	1.71	2.34 +/- 0.34	1.42	3.15	0.97	0.54	5.1
13616	1.84	2.74 +/- 0.74	1.41	3.97	0.96	0.67	5.6
14467	1.66	2.29 +/- 0.44	1.30	3.14	0.95	0.56	4.3
15318	1.55	2.11 +/- 0.26	1.23	2.84	0.98	0.59	4.9
16000	1.59	2.24 +/- 0.39	1.23	3.09	0.98	0.62	5.0
16169	1.60	2.27 +/- 0.47	1.27	3.19	0.97	0.65	4.9
17020	1.57	2.23 +/- 0.42	1.23	3.11	0.98	0.65	5.1
17871	1.51	2.13 +/- 0.38	1.17	2.96	0.98	0.63	5.0
18722	1.51	2.15 +/- 0.35	1.17	2.98	0.97	0.62	4.4
19573	1.52	2.16 +/- 0.38	1.16	3.01	0.97	0.62	4.3
20000	1.51	2.16 +/- 0.38	1.16	3.02	0.97	0.62	4.5