ACT β Humanoid Benchmark
ACT (Action Chunking Transformer, ResNet18 backbone, CVAE; Psi0 baselines/act) checkpoints trained on Humanoid Benchmark tasks.
Checkpoints
100 demonstrations per task. Every row is one training run; the epoch folders are its published checkpoints.
| Folder | Task | Robot | Cameras (model input) | Global batch | Proprio augmentation | Checkpoints |
|---|---|---|---|---|---|---|
b6-pedal-bin/ |
B6 β pick up a chip barrel, step on the pedal bin, drop it in | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-020 |
a3-fridge-open/ |
A3 β open the fridge door | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
a7-cabinet-close/ |
A7 β close the cabinet door | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
a10-trolley-push/ |
A10 β push a trolley | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
a20-basket-transfer/ |
A20 β transfer a basket | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
b7-black-clothes-hamper/ |
B7 β take the black garment off the hanger and drop it into the hamper | Unitree G1 (Inspire hands) | ego_view, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
a18-pour-water/ |
A18 β pour water from a pitcher into a glass | H2 (Sharpa hands) | head, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
a17-unscrew-bottle/ |
A17 β unscrew the cap of a bottle | H2 (Sharpa hands) | head, chest | 696 | earlier recipe (with delay/drop) | epoch-010, epoch-020 |
What an epoch folder contains
Every epoch-XXX/ folder is self-contained for evaluation and deployment: weights, model configuration, normalization
statistics and protocol.json (task, data, input/output dimensions, cameras, batch, steps, augmentation, file list).
Simple training records are included for reproducibility (training plan, training log, configs, metrics); optimizer /
scheduler / RNG / DeepSpeed resume state is not. code/ holds the adapter code used for training and inference
(code/VERSIONS.json: upstream commits and package versions).
Inputs and augmentation
Model inputs are identical across all benchmark models: G1 state 44-D (29 body joint angles + 6 + 6 measured hand values
- 3 projected gravity), H2 state 78-D (75 joint angles + 3 projected gravity), no joint velocities. Cameras: G1 ego view + chest; H2 head + chest (A14: head + chest + wrist). The third-person camera recorded in the datasets is for data review only and is never a model input. Each camera is a separate image input, except for Delta Flow and DiT4DiT, which place the views side by side in one frame. Training uses the Humanoid Benchmark online data augmentation spec v2 (15 photometric / geometric / sensor / lighting effects, one of them boosted to 1.5x per sample) with the model's own image augmentation turned off, plus joint-angle bias and noise on the proprioception (column "Proprio augmentation"; the earlier recipe also delayed the proprioception and dropped action-history rows).
A20 pre-reach trim v2, batch 512, step 4000 (2026-11-17)
Checkpoint: a20-basket-transfer/b512-trimv2/step-004000/.
Trained on 100 demonstrations (54,433 frames) from
humanoid-benchmark/a20-basket-transfer-pre-reach-trim-expert-v2, with batch size 512,
4,000 optimizer steps, seed 0, EMA weights (decay 0.9999), and Humanoid Benchmark
augmentation v2. These weights use the hb_baselines/1 format.
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). Normalization is stored in the weights.
The ResNet18 visual backbone was initialized with ImageNet weights and fine-tuned.
The checkpoint contains model.pt, config.json and protocol.json (weight size
and SHA-256). SHA256SUMS covers the checkpoint and the accompanying inference
source snapshot. The snapshot is under code/hb_baselines_v1/; its requirements
record the training environment. Optimizer/resume state is excluded.
After downloading this repository and installing its inference dependencies, load the checkpoint from the repository root:
import sys
sys.path.insert(0, 'code/hb_baselines_v1')
from baselines.runtime import BaselinePolicy
policy = BaselinePolicy(
'a20-basket-transfer/b512-trimv2/step-004000',
device='cpu', policy='act', robot='g1',
)
print(policy.info())
The saved weights passed SHA-256 verification and strict CPU loading before upload.
The offline metrics in config.json use training frames; they are not held-out
or closed-loop task success rates.
This release uses the accompanying benchmark baseline loader. The earlier
epoch-XXX releases retain their original Psi0 ACT loader and file format.
B6 batch 512, step 4000 (2026-11-17)
Checkpoint: b6-pedal-bin/b512/step-004000/. A new training run on B6; the earlier
B6 release b6-pedal-bin/epoch-020/ (Psi0 ACT loader, batch 696, earlier augmentation recipe) is kept unchanged.
Trained with evaluation/baselines/train.py --policy act --contract g1_sonic_dfx6 (Humanoid_benchmark commit
4385573e) on the first 100 episodes (65,127 frames) of the B6 v5t7 ready-start expert data, the data of the GR00T
b6-pedal-bin/b256 and Psi0 b6-pedal-bin/b512 checkpoints: batch size 512, 4,000 optimizer steps (about 31.4
epochs), AdamW lr 1e-4, 200-step linear warm-up then cosine decay to step 4000, seed 0, EMA weights (decay 0.9999),
bf16 autocast, ImageNet-initialized ResNet18 fine-tuned at 0.1x the learning rate (frozen BatchNorm), and Humanoid
Benchmark augmentation v2 (one call per sample on both views; joint-angle bias +-2 deg plus 0.25 deg noise on the 29
body joints only; no proprio delay or history drop; native random shift off). Before training, a real batch of 512
was checked for every model: state 44-D without velocity, only the body joints perturbed, action labels bit-identical
to the parquet files (training_records/batch_check_b6_v5.json). Final loss: L1 0.050, KL 0.006; offline check
(normalized action MSE on 256 training frames) 0.0054. These weights use the hb_baselines/1 format and the same
code/hb_baselines_v1/ loader as the A20 release above (byte-identical to the training code).
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). Normalization is stored in the
weights. 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 evaluation (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; success = the barrel inside the bin and
released): 8/50 (Wilson 95 % 8-29 %); the barrel touched 45, lifted 28, pedal pressed 10, in the bin 10. Same
seeds and protocol: GR00T b6-pedal-bin/b256/step-004000 39/50, Psi0 b6-pedal-bin/b512/step-004000 30/50,
FM-DiT 0/50, DP 0/50 (trained with this recipe).
Files: model.pt, config.json, protocol.json (weight size and SHA-256), SHA256SUMS (the checkpoint, the
loader snapshot and the training records) and training_records/ (run configuration run.json, train_log.jsonl,
train.log, the pre-training batch check). Optimizer/resume state is excluded. Load it as the A20 example above with
'b6-pedal-bin/b512/step-004000'.