Venue: CoRL 2025 · arXiv: 2509.07962 Category: VLA Architecture Trend tag: Force as first-class policy signal

flowchart LR
  V[Vision] --> VLM[VLM]
  L[Language] --> VLM
  VLM --> DEC[Action decoder]
  T[Joint torque history] --> TOK[Summarize as<br/>SINGLE token]
  TOK --> DEC
  DEC --> A[Actions]
  DEC --> TP[Predicted torque<br/>aux objective]

Most VLAs are vision-and-language only. Contact-rich tasks (wiping, insertion, pushing doors) quietly fail because the model can't tell success from failure until it sees a visual cue — far too late. Joint torque is the natural early signal, but naive injection (e.g., per-timestep torque tokens) destabilizes training and regresses non-contact tasks.

Systematic study of where and how to inject torque into a pretrained VLA (base model: π₀, a flow-matching VLA). Two key findings:

1. Where: place torque in the decoder as a single history-summary token (not per-timestep, not in the encoder) — this "preserves the original input pattern of the decoder" and balances informativeness with architectural stability.
2. How (auxiliary objective): also predict torque as an auxiliary output via a joint action–torque diffusion loss (ℒ_joint = ℒ_action + β·ℒ_torque), inspired by joint prediction-and-planning in autonomous driving. This pushes the model to build a physically grounded internal representation and further improves performance.

On 5 contact-rich tasks (20 trials each, π₀ baseline → TA-VLA): Button Pushing 5/20→18/20, Charger Plugging 0/20→17/20, USB Plugging 0/20→17/20, Socket Unplugging 16/20→19/20, Door Handle Turning 2/20→15/20. Raises the typically near-zero contact-rich success rates to ~75–95% with essentially no regression on the 5 regular (non-contact) tasks. Baselines include ACT, RDT-1B, and π₀.

Part of CoRL 2025's force-as-first-class trend alongside UniFP (Best Paper), DexSkin, and Tactile Beyond Pixels. Delivers a clean, composable recipe that any existing VLA can adopt.

• arXiv: https://arxiv.org/abs/2509.07962

• UniFP
• Survey (CoRL 2025)

← Back to CoRL-2025