GR00T-N1.5-robocasa-baseline

GR00T-N1.5 finetuned on RoboCasa Kitchen. This is the reference (baseline, non-MoE) policy used throughout the quantizability gate project, and the policy that every RoboCasa gate module in this collection was measured against.

• Benchmark: RoboCasa Kitchen, 24 tasks
• Embodiment: RoboCasa mobile manipulator. Embodiment tag in the shipped metadata is new_embodiment.
• Observations: three 256x256 RGB views — left_view, right_view, wrist_view — at 20 fps
• State: base position (3) + base rotation (quaternion, 4) + absolute and relative end-effector position (3+3) and rotation (quaternion, 4+4), plus gripper
• Actions: delta end-effector control with a gripper channel, action horizon 16
• Finetuning data: MimicGen demonstrations, LeRobot dataset kimtaey/robocasa_mg_gr00t_300
• Checkpoint: step 60,000 = 1.85 epochs over the dataset (the run was configured for 2 epochs / 60k steps and 60k came first). Final training loss 0.0324, from 0.577 at step 0.

Measured closed-loop performance

24 tasks x 50 episodes, 4 denoising steps:

configuration	success rate	steps on successful episodes
this policy, uncompressed	0.657	327.0
blanket K=2 action-chunk compression	0.598	214.0

Those two rows are the endpoints of the trade-off that the gate modules in this collection are trying to beat; see for example prehj/quantgate-student-robocasa-phase5-softB.

Architecture

GR00T-N1.5 (model_type: gr00t_n1_5, architectures: ["GR00T_N1_5"]), unchanged from the NVIDIA release:

VLM backbone	Eagle — NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops
features tapped at	select_layer: 12, backbone embedding dim 2048
backbone tuning	tune_llm: false, tune_visual: true (vision tower trained, LLM frozen)
action head	flow-matching DiT, 16 layers, 32 heads, head dim 48, cross-attn dim 2048, interleaved self-attention, ada_norm, dropout 0.2
VL self-attention	4 layers, 32 heads, head dim 64
action head tuning	tune_diffusion_model: true, tune_projector: true
action space	action_dim: 32 (padded), max_state_dim: 64
action horizon	16
inference denoising steps	4
dtypes	model_dtype: float32, compute_dtype/torch_dtype: bfloat16

Training recipe

Read directly out of this checkpoint's trainer_state.json and the run's training_args.bin:

steps	60,000 (this is the final checkpoint; max_steps == global_step)
batch size	64 effective — 32 per device x 2 GPUs, no gradient accumulation
learning rate	1e-4, cosine schedule, warmup ratio 0.05
weight decay	1e-5
precision	bf16
epochs	1.85 (60k steps reached first; num_train_epochs: 2)
dataloader workers	32
final loss	0.0324

How to load and serve

Requires NVIDIA Isaac-GR00T and transformers==4.51.3 (the version this checkpoint was written with).

huggingface-cli download prehj/GR00T-N1.5-robocasa-baseline --local-dir ./ckpt

from gr00t.model.policy import Gr00tPolicy
from gr00t.experiment.data_config import DATA_CONFIG_MAP

data_config = DATA_CONFIG_MAP["robocasa_panda_omron"]
policy = Gr00tPolicy(
    model_path="./ckpt",
    modality_config=data_config.modality_config(),
    modality_transform=data_config.transform(),
    embodiment_tag="new_embodiment",
    device="cuda",
)
action_chunk = policy.get_action(observation)   # 16 steps

Or serve it over the Isaac-GR00T inference service and drive it from the simulator client:

python scripts/inference_service.py --server \
  --model_path ./ckpt --embodiment_tag new_embodiment --denoising_steps 4

experiment_cfg/metadata.json ships in this repo and carries the state/action normalisation statistics and the modality schema — the policy will not produce correct actions without it, so keep the directory layout intact.

Files

file	what it is
model-0000{1,2}-of-00002.safetensors, model.safetensors.index.json	the weights (~7.1 GB)
config.json	GR00T-N1.5 architecture config
experiment_cfg/metadata.json	embodiment schema + normalisation statistics
trainer_state.json	full 60k-step loss history