GR00T-N1.5 β€” RoboCasa Kitchen baseline policy

GR00T-N1.5-3B finetuned on RoboCasa Kitchen (robocasa_mg_gr00t_300, 24 tasks), batch size 64, 60k steps. This is the uncompressed baseline in the quantizability-gate work β€” the policy that produces the 16-step action chunks which the gate then decides whether to temporally compress.

Closed-loop, 24 tasks Γ— 50 episodes each:

policy success avg steps
this model, uncompressed 0.657 327
+ naive K2 (compress every chunk) 0.598 221
+ Aβ€² gate (gemma4 teacher), Ο„=0.5 0.667 258

Benchmark reproducibility noise is Β±1.5 pp.

Which checkpoint do I need?

Usage

from gr00t.model.policy import Gr00tPolicy

policy = Gr00tPolicy(model_path="prehj/groot-n15-robocasa-kitchen-baseline",
                     embodiment_tag="new_embodiment", denoising_steps=4)
chunk = policy.get_action(obs)["action_pred"]     # (16, 12)

Then gate and merge:

from gate_api import QuantizabilityGate          # from the A' repo
from merge_k2 import merge_k2

gate = QuantizabilityGate("checkpoints/A_robocasa_gemma4.pt",
                          "checkpoints/robocasa_task_embeddings.npz")
if gate(views, instruction)["confidence"] >= tau:
    chunk = merge_k2(chunk, delta_dims=range(5, 11), state_dims=[11])

Ο„ is an operating point to search or learn, not a constant β€” see the Aβ€² card.

Action space (12 dims)

0–4 unused (always zero) Β· 5–7 EE delta xyz Β· 8–10 rotation Β· 11 gripper (0/1). Under K2, dims 5–10 sum across the merged pair and dim 11 takes the last value.

Running the benchmark

Environment setup, the three-process runtime layout, the full task list, how to sweep tau, and the pitfalls that have actually cost us runs are written up here: RoboCasa evaluation setup guide

Related

Code and evaluation harness: https://github.com/rakybond007/GR00T-action-quantization/tree/action-quantization-gate-v2

Downloads last month
7
Safetensors
Model size
3B params
Tensor type
F32
Β·
BF16
Β·
Video Preview
loading