SimpleTouch: Can Vision-Language-Action Models Master Contact-Rich Manipulation Without Tactile Policy Pretraining?

Can a pretrained vision-language-action model learn strong contact-rich manipulation from limited task demonstrations when given a pretrained tactile encoder, without additional tactile policy pretraining or a separate visuotactile alignment stage?

Contact-rich manipulation states are often visually ambiguous, yet adding tactile observations to pretrained VLA models is difficult because touch is outside their pretraining modalities. Directly injecting touch during task fine-tuning frequently gives only limited gains or even reduces success, while existing remedies require large-scale tactile policy pretraining or separate representation alignment, adding data and training overhead.

Prior methods either pretrain tactile policies on large heterogeneous manipulation datasets, as in FTP-1, or align visual and tactile representations before policy learning, as in VITaL; task-level baselines include ACT, UniVTAC-ACT and Tactile-VLA. These approaches front-load data collection and training stages, and naive injection of tactile features into a pretrained VLA does not reliably improve success.

SimpleTouch augments π0.5 with a tactile expert that consumes all current tokens from a frozen T3 tactile encoder and exposes them to the action expert via layerwise attention. Training uses only 50 demonstrations per task and jointly optimizes the π0.5 visual-language backbone, action expert, and tactile expert, keeping the tactile encoder frozen. The tactile expert is additionally supervised by multi-horizon prediction of future pooled tactile latents via flow matching; at deployment only the cached current-touch representation is used, so no future tactile prediction is needed.

On six UniVTAC tasks, SimpleTouch ranks first on every task and averages 77.5% success versus 66.7% for pretrained FTP-1, 45.2% for FTP-π0.5, 46.3% for π0.5, and 30.3% for ACT. Excluding two visually dominated lift tasks, SimpleTouch averages 74.3% versus 59.5% for FTP-1. On four real-robot tasks it averages 71.3%, compared with 62.5% for FTP-1, with gains of 20 points on Wipe Board and 10 points on Pick Up Chips; Insert USB remains 30% for both. Ablations show full 197-token tactile context averages 74.3% versus 39.0% vision-only and 57.5% CLS-only; trainable T3 drops to 70.5%. Multi-horizon future prediction gives 74.3% versus 68.8% without it and 68.5% single-horizon. At 20 demonstrations SimpleTouch reaches 38% and 57% on Put Bottle and Insert Hole, versus 19% and 47% for FTP-π0.5. Inference in the Fast formulation averages 0.623 s latency with 74.3% success, versus 1.020 s/70.3% for Joint and 1.463 s/67.3% for IDM.

The authors state Insert USB remains challenging, with only 30% success for both SimpleTouch and FTP-1, and low-data absolute performance is modest, e.g. 23–43% at 10 demonstrations. Broader limitations are not explicitly discussed, but the real-robot evaluation covers four tasks on a single platform with 20 trials per task, generalization beyond the tested contact tasks and hardware is unproven, and code availability is not reported.