model card

README.md

@@ -0,0 +1,61 @@
+---
+license: apache-2.0
+base_model: nvidia/GR00T-N1.6-3B
+tags:
+- robotics
+- gr00t
+- unitree-g1
+- inspire-hand
+- manipulation
+datasets:
+- birbirll/g1-inspire-piston-pick-place
+---
+
+# GR00T N1.6 — G1 + Inspire piston pick-and-place
+
+Fine-tune of `nvidia/GR00T-N1.6-3B` on
+[birbirll/g1-inspire-piston-pick-place](https://huggingface.co/datasets/birbirll/g1-inspire-piston-pick-place)
+(102 success episodes, LeRobot v2.1). Picks up a piston from a table with the
+right Inspire hand (Unitree G1, fixed base). Closed-loop verified in IsaacLab.
+
+## Contents
+
+- Model weights (bf16 safetensors) + processor/experiment configs — loadable
+  with Isaac-GR00T's `run_gr00t_server.py` (`--embodiment_tag NEW_EMBODIMENT`).
+- **`g1_inspire_modality_config.py`** — the modality config used for training
+  (register via `--modality-config-path`): state = arms 14 + hands 12 + waist 3
+  (dims 29:63 of the raw state are unused tactile); action = 30-D
+  (left_arm 7 | right_arm 7 | left_hand 6 | right_hand 6 | base_height 1 |
+  navigate 3), 30-step horizon; arms trained RELATIVE (decoded to absolute by
+  the server).
+
+## Training recipe (single RTX 4090, ~2.5 h)
+
+```bash
+python gr00t/experiment/launch_finetune.py \
+  --base-model-path nvidia/GR00T-N1.6-3B \
+  --dataset-path <local dataset> \
+  --embodiment-tag NEW_EMBODIMENT \
+  --modality-config-path g1_inspire_modality_config.py \
+  --num-gpus 1 --output-dir ./out \
+  --max-steps 10000 --save-steps 1000 --save-total-limit 2 \
+  --global-batch-size 8 --gradient-accumulation-steps 4 \
+  --state-dropout-prob 0.8 \
+  --color-jitter-params brightness 0.3 contrast 0.4 saturation 0.5 hue 0.08
+```
+
+Key knobs: `--state-dropout-prob 0.8` (forces vision conditioning — without it
+the policy shortcuts through proprioception and ignores the camera);
+effective batch 32; default LR 1e-4; loss ~1.16 → ~0.011. On 24 GB GPUs you
+must additionally set the optimizer to `paged_adamw_8bit` and enable gradient
+checkpointing in `launch_finetune.py` (upstream defaults OOM); on A100-class
+hardware the upstream defaults are fine.
+
+## Serving notes
+
+- The processor expects `shortest_image_edge: 256` — send **raw camera
+  frames** and let the server's own transform do all resizing. A client-side
+  resize (e.g. to 224) silently degrades the policy.
+- Deployment chunk pacing: 30-step chunks; we execute 20 steps per inference
+  cycle, **temporal ensembling OFF** (at low replan rates ACT-style ensembling
+  averages stale chunks — measured grasp closure 0.22 with vs 0.99 without).