DP β 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/ (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
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
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.