References & Resources

This Project

• Code: github.com/bushuyeu/smolvla-force-torque
• Checkpoints: huggingface.co/bushuyeu/smolvla-ft-checkpoints (7.4GB compressed)

WandB Dashboards

• Experiment 1 (Force Token): wandb.ai/…/smolvla-force-token
• Experiment 2 (Force Distill): wandb.ai/…/smolvla-force-distill
• Experiment 4 (Dual-Rate): wandb.ai/…/smolvla-dual-rate

Base Model

• SmolVLA paper: arXiv:2506.01844
• SmolVLA weights: huggingface.co/lerobot/smolvla_base
• LeRobot framework: github.com/huggingface/lerobot

Dataset

• RH20T (Robot Handling 20 Tasks): huggingface.co/datasets/hainh22/rh20t
  • 110K sequences across UR5, Franka, KUKA, Flexiv
  • Synchronized 6-axis F/T, joint torques, RGB-D, audio
  • We used the UR5 subset (30 episodes, 4,933 frames)

Key Papers

Force-Augmented VLAs

• ForceVLA (May 2025) – Force-aware MoE fusion on pi0
• FD-VLA (Feb 2026) – Force distillation from vision on pi0
• FTACT (Sep 2025) – F/T-aware ACT
• FoAR (2025) – Force-aware residual policy (inspiration for Exp 4)
• CompliantVLA-adaptor (Jan 2026) – VLM-guided impedance control

Architecture Design

• TA-VLA (2025) – Torque in decoder, not encoder
• Direction Matters (2025) – Directional F/T encoding

Tactile Integration

• VLA-Touch (Jul 2025) – Dual-level tactile adapter
• TaF-VLA (2025) – Combined tactile + F/T

See the Related Work page for a full landscape analysis.

Intrinsic AI Challenge

• Challenge site: intrinsic.ai/challenge
• Task: Cable insertion on UR5e + Axia80 F/T sensor + 3 Basler cameras + Robotiq Hand-E
• Prize pool: $180K
• Timeline: March-August 2026

Citation

@misc{bushuyeu2026smolvla-ft,
  title={SmolVLA Force-Torque Experiments: Adding Touch Sensing to a 0.45B VLA},
  author={Bushuyeu, Pavel},
  year={2026},
  url={https://github.com/bushuyeu/smolvla-force-torque}
}