AIAny
Icon for item

$N_0$-VTLA: Scaling Vision-Tactile-Language-Action Model with Latent Tactile Tokens

Enables tactile-aware robot manipulation by pretraining a vision–tactile–language–action foundation model and improving offline policies with ALTER. Combines large-scale NeoData visuo-tactile pretraining, a latent tactile pathway for predictive touch signals, and advantage‑conditioned offline RL for contact-rich tasks.

Introduction

Why this matters Contact-rich manipulation remains a major bottleneck for general-purpose robot skills: visual cues alone often miss contact subtlety, while tactile signals are sparse and hardware-dependent. This work demonstrates a practical recipe to transfer large-scale visuo-tactile priors into downstream policies and to improve deployed policies from logged experience, producing measurable gains on both simulated and real-robot benchmarks.

Key Findings
  • Pretraining on NeoData yields transferable contact priors that the policy can distill into fine motion control via a dedicated latent tactile pathway; this staged integration recovers tactile utility without retraining entire vision backbones.
  • ALTER, an advantage‑conditioned offline RL procedure, turns relative progress and event labels into binary advantage conditions, enabling policy improvement from fixed deployment logs and human corrections.
  • Empirical gains are substantial: the full system wins all nine NeoReal real-robot tasks, reaches 63.8% mean success on a 20-task simulated suite versus 44.0% for the strongest baseline, and—when combined with ALTER—achieves 75–95% success on three long-horizon real-robot tasks.
Who It's For and Tradeoffs

Great fit if you work on contact-sensitive robot manipulation and can collect or access tactile-enabled datasets and hardware. The approach is valuable for insertion, deformable object handling, and dual-arm contact tasks that require reflexive touch feedback. Look elsewhere if you lack tactile sensors or deployment logs: benefits rely on large visuo-tactile corpora (NeoData), staged tactile integration, and sufficient compute to train/finetune the latent tactile pathway. There are also engineering costs for tactile sensing, data labeling of contact events, and offline-data curation for ALTER.

Where It Fits

Positions itself between purely vision-based policy backbones and end-to-end tactile specialists: it leverages vision-scale architectures while adding a lightweight, predictive tactile stream and an offline RL refinement step, making tactile capability more modular and more amenable to real-world deployment.

Information

  • Websitearxiv.org
  • OrganizationsNeoteAI Team, Fudan TEAI Team
  • Published date2026/07/26

More Items

Regularizes latent world models by replacing the Epps–Pulley Gaussianization objective with a quantile–quantile matching loss that aligns projected latent samples to rank-matched Gaussian quantiles, improving tail correction and planning success via cross-batch ranking.

Transforms open-ended LLM optimization into self-verifiable reinforcement learning by turning tasks into proxy environments that produce deterministic, rule-based rewards. Proposes RLSVR and SpyRL — an information-asymmetric self-play scheme where agents vote to identify a preassigned spy, yielding verifiable rewards without human annotation. Demonstrated on summarization, creative writing and mathematical reasoning.

Allocates token-level credit in rubric-conditioned GRPO by counterfactually replaying the same response under rubric and criteria-free prompts, using tokenwise log-likelihood contrasts to compute bounded, response-normalized weights that redistribute GRPO advantages without training an auxiliary scorer.