Downloads · 30 days
15
32% of all-time downloads
prehj/GR00T-N1.5-robocasa-atq-v22so3-2x-pool-attention-60k-20260921
GR00T-N1.5-robocasa-atq-v22so3-2x-pool-attention-60k-20260921 is a robotics model from prehj. Use it for the robotics task on the model card, and read the license before you ship it in a product.
RoboCasa ATQ v22, SO(3) merged rotation targets, 2x compression, four experts. Trained from the base GR00T model for 60,000 steps with global batch 64, seed 42. Pooling: attention, shared by router and confidence head…
Downloads · 30 days
15
32% of all-time downloads
All-time downloads
47
Public
Parameters
2.8B
8.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8.1 GB · 100%
How the weights are stored.
BF161.7B · 58%
From the Hugging Face model README
RoboCasa ATQ v22, SO(3) merged rotation targets, 2x compression, four experts. Trained from the base GR00T model for 60,000 steps with global batch 64, seed 42. Pooling: attention, shared by router and confidence head; state features still use mean pooling. Router KL coefficient: 0.1. Balance coefficient: 0.05. Confidence regression remains enabled (coefficient 0.1). Attention uses the existing single-query four-head module with no additional tuning. Last-token uses the last unpadded VL position. The conf input remains detached as in the baseline. Both original router regularizers remain enabled. Loading last-token checkpoints requires the accompanying last-token support in flow_matching_action_head.py; use the matching GR00T-action-quantization repository. Source changes and recipe are included under pooling_implementation/. Final inference weights and normalization metadata only; optimizer states omitted. Not yet evaluated. Full settings and weight hashes: ablation.json.