g1-inspire-piston-egoonly-22170

EgoVLA — G1 Inspire piston/pipette, EGO-CAMERA-ONLY (22170 steps)

Single-view (ego camera only) EgoVLA checkpoint for the bimanual G1 piston/pipette task. This is our best-performing checkpoint to date.

What's in here

config.yaml                     # required — the loader reads it from the ckpt's grandparent dir
dataset_statistics.json         # required — action/state un-normalization
final_model/pytorch_model.pt    # 3.8 GB

Keep the directory structure intact: read_mode_config resolves the run dir as checkpoint_pt.parents[1], so config.yaml and dataset_statistics.json must sit one level above final_model/.

Model

• run_id: starvla_egovla_piston_bir_dart_joint — 22170 steps (is_resume: true)
• EgoVLA (VILA: SigLIP-384 + Qwen2-1.5B) + MLP L1 action head
• Action: 26-dim absolute joint targets [left_arm 7 | right_arm 7 | left_hand 6 | right_hand 6] (Inspire-6 hands)
• action_horizon 30, obs_image_size 384×384, state_dropout_prob 0.8
• One camera (ego_view) → one <image> token
• Data mix: unitree_g1_piston_bir_dart_pnp (67 clean demos + 16 DART recovery episodes)

Before you load it

• base_vlm in config.yaml points at /storage/scratch1/0/hding95/ckpts/ego_vla_checkpoint/ckpt-6720. Repoint it to wherever your copy of the EgoVLA base VLM lives.
• data_root_dir likewise points at our scratch space. Only matters if you resume training.
• The data mix name must be registered in examples/.../data_registry/data_config.py for the norm processor to resolve robot_type.

Eval (10 episodes, our IsaacLab piston stack, inference_hz 1.25)

S3/S5 are judged on the physically measured plunger travel, not on whether a press was commanded. Commanded-press rates are 100%/100% — the gap at S5 is that the second press is physically shallower (median ≈ 4.6 mm vs ≈ 7.1 mm in the demos), and the pass mark is 4 mm. Because that distribution straddles the threshold, single 10-episode S5 numbers are noisy: two runs of an identical serving config gave 9/10 and 4/10. Compare over ≥ 20 episodes.

Note on the press loss weighting

This checkpoint already trains with grasp_loss_weight: 3.0 on grasp_weight_dim: 25 (= action[25], the right thumb_yaw = the plunger press channel), grasp_weight_thresh: -0.143 — i.e. the same 3× press weighting as the multi-view runs. It presses reliably. So a multi-view checkpoint that fails to press is not failing because of that setting.

Serving

Served through VLAPolicyBridge with a GR00T-protocol ZMQ server; the policy is view-agnostic (26-dim joint chunk out). Recommended bridge config: temporal ensembling ON with the hand groups excluded from blending (--ensemble_no_blend left_hand,right_hand) — blending rounds off the press pulse, which is impact-driven.