HumanoidBench GR00T models

GR00T models trained for the HumanoidBench tasks in latency-sensitive-bench. The repository layout is:

balance_simple/
├── checkpoint-10000/  # published
└── checkpoint-15000/  # published automatically after the future 15k run

The 15k run has not been started yet. Its fixed W&B project is finetune-gr00t-n1d7, with run id and name g1_balance_simple_obs198_h40_repeat_fixedz0_onestep_dropout0_15k.

Balance Simple

balance_simple/checkpoint-10000 is the completed 10k model from the formal g1_sonic_balance_simple recipe. The future balance_simple/checkpoint-15000 will use the same recipe with a fresh 15k schedule:

• input: 198D state and 256×256 head-camera RGB image;
• output: H40 repeated-current 43D absolute action;
• training: fixed Gaussian seed 0, one-step flow, action-head dropout 0, global batch 128, and 250 warmup steps; checkpoint-10000 used 10,000 optimizer steps, while the future checkpoint-15000 uses 15,000 steps.

Download the completed 10k inference bundle before passing its local path to GR00T:

third_party/Isaac-GR00T/.venv/bin/hf download \
  latency-sensitive-bench/humanoidbench \
  --include 'balance_simple/checkpoint-10000/**' \
  --local-dir outputs/official_gr00t/models/humanoidbench

After the future 15k run completes, replace checkpoint-10000 with checkpoint-15000 to download its automatically published bundle.

The published bundle contains model safetensors, model/processor config, embodiment mapping, and statistics. DeepSpeed optimizer state, RNG state, scheduler state, and W&B artifacts are intentionally not published; this bundle supports inference, not optimizer resume.

The base model is nvidia/GR00T-N1.7-3B. No additional license metadata is asserted here.