armteam/crab-manipulation-dataset-labeled

Crab Manipulation Dataset — Reward Labeled

Same demonstrations as crab-manipulation-dataset, augmented with per-episode reward parameters for Safety-Aware Reward-Weighted Flow Matching (SA-RWFM) training.

What's Different

Each session's meta/ folder contains an additional reward_params.json with:

• —Tactile analysis: contact fractions, force statistics, damage flags
• —Gripper analysis: open/close thresholds, velocity analysis
• —Reward parameters: success/drop/damage rewards, risk weighting, step-level scores
• —Fragility labels: whether the object is fragile (soft) or not

Dataset Structure

Clean demonstrations (27 sessions)

Negative examples (7 sessions, damage_flag=true)

Reward Label Format

{
  "tactile_scale": 1695.0,
  "tactile_active": true,
  "fragile": true,
  "damage_flag": false,
  "contact": { "contact_frac_L": 0.20, "contact_frac_R": 0.27, ... },
  "gripper_left": { ... },
  "gripper_right": { ... },
  "params": { "R_succ": 15.0, "R_drop": 20.0, "R_damage": 20.0, ... }
}

Related

• —crab-manipulation-dataset — Same data without reward labels
• —crab-smolvla-rwfm — SA-RWFM teacher trained on this data
• —crab-smolvla-hapticsvla — HapticVLA distilled from the teacher

Citation

If you use this model, please cite our paper:

@article{gubernatorov2026hapticvla,
  title={HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing},
  author={Gubernatorov, Konstantin and Sannikov, Mikhail and Mikhalchuk, Ilya and Kuznetsov, Egor and Artemov, Makar and Ouwatobi, Ogunwoye Faith and Fernando, Marcelino and Asanov, Artem and Guo, Ziang and Tsetserukou, Dzmitry},
  journal={arXiv preprint arXiv:2603.15257},
  year={2026}
}