File size: 4,876 Bytes
ab621ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
---
tags:
- robotics
- humanoid
- imitation-learning
- diffusion-policy
---
# DP — Humanoid Benchmark

Diffusion Policy (DDPM) checkpoints trained on Humanoid Benchmark tasks. The policy uses a ResNet18 image encoder and
an AdaLN DiT action backbone (hidden dimension 384, 4 layers, 8 attention heads), the same networks as the FM-DiT
baseline (`humanoid-benchmark/fm-dit`); only the generative objective differs: epsilon-prediction DDPM with 100
training timesteps (`squaredcos_cap_v2` schedule, clipped samples), sampled over all 100 timesteps.

## B6 pedal bin, batch 512, step 4000, with resume state (2026-11-17)

Checkpoint: [`b6-pedal-bin/b512/step-004000/`](b6-pedal-bin/b512/step-004000) (the run's last checkpoint).

Trained with `evaluation/baselines/train.py --policy dp --contract g1_sonic_dfx6` (Humanoid Benchmark commit
`4385573e`; `code/hb_baselines_v1/` is byte-identical to its `evaluation/baselines` and `evaluation/augmentation`, and to
the snapshot in `humanoid-benchmark/fm-dit`) on the first 100 episodes (65,127 frames, about 31.4 epochs) of the B6 v5t7
ready-start expert data, the data of the GR00T `b6-pedal-bin/b256`, Psi0 `b6-pedal-bin/b512`, FM-DiT and ACT
`b6-pedal-bin/b512` checkpoints (`humanoid-benchmark/b6-pedal-bin` `v5/`, episodes 0-99). `hb_baselines/1` format,
22.5 M parameters (11.2 M in the ResNet18). Training: batch size 512 (fits the 32 GB RTX 5090), 4,000 optimizer steps,
seed 0, no held-out episodes, Humanoid Benchmark augmentation v2 (`hb-aug-v2-20261109`; joint-angle bias +-2 deg plus
0.25 deg noise on the 29 body joints only), AdamW (lr 1e-4, betas 0.95/0.999, weight decay 1e-6), the
ImageNet-initialized ResNet18 fine-tuned at 0.1 x the learning rate with frozen BatchNorm, 200-step linear warm-up then
cosine decay to 0 at step 4,000, bf16 autocast, EMA weights (0.998 at step 4,000). Before training a real batch of 512
was checked: state 44-D without velocity, only the body joints perturbed, action labels bit-identical to the parquet
files (`batch_check_b6_v5.json`). Final DDPM loss 0.049; offline check (normalized action MSE on 256 training frames)
0.034 at step 2,000 and 0.030 at step 4,000.

Inputs: Unitree G1 state44 and separate `ego_view` and `chest` RGB images, resized to 168 x 224. Output: 16 actions of
dimension 76 (64 SONIC motion-token values and 6 measured values for each INSPIRE hand). Training instruction: "Pick up
the chip barrel from the cart, walk to the pedal bin, step on its pedal and drop it in."

Closed-loop result (Humanoid Benchmark `debug/newton/b6_policy_eval.py --protocol v5_ready_start` with the v5 task
configuration; 50 unseen layouts, seeds 180000-180049, 45 policy seconds): **0/50**; the robot never touched the
barrel. Same seeds and protocol: GR00T `b6-pedal-bin/b256/step-004000` 39/50, Psi0 `b6-pedal-bin/b512/step-004000`
30/50, ACT (`humanoid-benchmark/act` `b6-pedal-bin/b512/step-004000`, this recipe) 8/50, FM-DiT
(`humanoid-benchmark/fm-dit` `b6-pedal-bin/b512/step-004000`, this recipe) 0/50.

Files under `b6-pedal-bin/b512/`: `step-004000/` (`model.pt` EMA weights, `config.json` with the full configuration,
normalization statistics and the offline metrics, `protocol.json` with the weight size and SHA-256), `train_state.pt`
(resume state at step 4,000: raw policy weights, AdamW, LR scheduler, EMA, step, configuration; its EMA equals
`step-004000/model.pt`), `run.json` (command line, configuration, augmentation preflight), `train_log.jsonl`,
`train.log`, `batch_check_b6_v5.json`, and `SHA256SUMS` (these files and the source snapshot; `sha256sum -c
b6-pedal-bin/b512/SHA256SUMS` from the repository root). Local paths in the training records are shortened to `<data>/`
and `~/`.

### Load for inference

```python
import sys
sys.path.insert(0, 'code/hb_baselines_v1')
from baselines.runtime import BaselinePolicy

policy = BaselinePolicy(
    'b6-pedal-bin/b512/step-004000',
    device='cpu', policy='dp', robot='g1',
)
print(policy.info())
```

### Resume training

```bash
hf download humanoid-benchmark/dp --include 'b6-pedal-bin/b512/*' --include 'code/*' --local-dir dp
hf download humanoid-benchmark/b6-pedal-bin --repo-type dataset --include 'v5/*' --local-dir b6-pedal-bin
mkdir -p runs/b6_dp && cp dp/b6-pedal-bin/b512/train_state.pt runs/b6_dp/
python dp/code/hb_baselines_v1/baselines/train.py --policy dp --contract g1_sonic_dfx6 \
    --dataset b6-pedal-bin/v5 --episodes 0:100 --out runs/b6_dp --resume \
    --steps 8000 --batch-size 512 --save-every 2000 --device cuda:0
```

`--resume` reads `<out>/train_state.pt` and continues from step 4,000 (check for `{"event": "resumed", "step": 4000}` in
the log; without that file the run starts from step 0). Keep `--policy dp`, `--contract g1_sonic_dfx6` and the default
model and augmentation settings; `--steps` is the new total. The resume itself was not test-run from this layout.