gr00t_real_deploy_official

UR5 + RealSense + FLIR + Robotiq deploy / inference code for fine-tuned GR00T-N1.7 policies. No checkpoints in this repo — get them from yqi19/gr00t_real_checkpoint.

Branches

Branch	What it is
correct-version	Eval code with all four post-state-B fixes applied: _build_gr00t_obs passes native-res RGB, action_ts anchored at time.time() + 0.1, j_init reads j_init_deg from config, _IMG_W/_IMG_H removed.
smooth-version	Pre-fix snapshot — all four bugs intact: extra cv2.cvtColor(BGR2RGB) channel flip, cv2.resize to 224×224 (defeats LetterBox), action_ts anchored at obs_ts[-1] (chunk all stale), hardcoded j_init.
latency-version	correct-version + two new CLI flags for tuning policy execution latency: --action_exec_latency (default 0.1s, forward buffer) and --action_start_step (default 0, skip first N predicted steps).

Quick reference

# correct (default behavior)
git clone -b correct-version https://huggingface.co/yqi19/gr00t_real_deploy_official deploy_correct

# latency (with tunable flags)
git clone -b latency-version https://huggingface.co/yqi19/gr00t_real_deploy_official deploy_latency
cd deploy_latency
# example: skip first 2 chunk steps, execute next 6
python eval/eval_gr00t.py \
    --checkpoint /path/to/checkpoint \
    --task "pick up bread" \
    --action_start_step 2 \
    --steps_per_inference 6 \
    --action_exec_latency 0.1

See BACKUP_NOTE.md on each branch for behaviour details.