HapticVLA: Tactile Skill Without Touch Sensors at Inference

Skoltech researchers created the HapticVLA VLA model, which learns gentle contact manipulation using tactile sensors only during training and operates without them, achieving an average success rate of 86.7% on fragile objects in real-world experiments.

Compiled by Sergey KostenchukPublished 2026-09-29Updated 2026-09-30

2026-09-29 Research

🤖 HapticVLA: How to teach a robot to feel without touch sensors?

The Intelligent Space Robotics Laboratory (Skoltech) published HapticVLA, accepted to IROS 2026: a VLA model learns gentle contact manipulation, and tactile sensors are only needed during training; at inference, it uses vision and state.

🌍 Tactile sensors for robots are expensive and poorly transferable between platforms. Here, the sense of touch is distilled into the model's weights: during training, the tactile signal serves as a reward and teacher, and is not needed at inference.

👤 On fragile objects, the average success rate is 86.7% — higher than baseline VLAs, even without sensors at inference. The code (a LeRobot fork with a Crab stand of two SO-101s), weights, and datasets are open: it can be reproduced without buying sensors. The work was done by master's students from Yandex and Skoltech's AI Robotics program in six months.

Source 1: https://arxiv.org/abs/2603.15257 Source 2: https://advanced-robotic-manipulation.github.io/websites/hapticvla/

Sources