ARM·ISR Lab / Skoltech   HapticVLA
Preprint · Under review · 2026

HapticVLA

Contact-rich manipulation via a Vision-Language-Action model—without inference-time tactile sensing.

Konstantin Gubernatorov*, Mikhail Sannikov*, Ilya Mikhalchuk*, Marcelino Fernando, Egor Kuznetsov, Ogunwoye Faith Ouwatobi, Artem Asanov, Makar Artemov, Ziang Guo, Dzmitry Tsetserukou * Equal contribution · Intelligent Space Robotics Laboratory, Skoltech
01

Abstract

preprint · under review

Tactile sensing is a crucial capability for Vision-Language-Action architectures, as it enables dexterous and safe manipulation in contact-rich tasks. However, reliance on dedicated tactile hardware increases cost and reduces reproducibility across robotic platforms.

We argue that tactile-aware manipulation can be learned offline and deployed without direct haptic feedback at inference. HapticVLA proceeds in two tightly-coupled stages. Safety-Aware Reward-Weighted Flow Matching (SA-RWFM) trains a flow-matching action expert on precomputed tactile safety rewards that penalize excessive grasping force and suboptimal trajectories. Tactile Distillation then transfers this tactile-aware capability into a standard VLA by distilling a compact tactile token from the SA-RWFM teacher and training a student that predicts it from vision and proprioception alone.

On real-world experiments, HapticVLA achieves a mean success rate of 86.7%, consistently outperforming baseline VLAs — including versions provided with direct tactile feedback during inference.

Success rate — HapticVLA vs. baselines
SmolVLA SmolVLA w/ SA-RWFM X-VLA VLA-0 HapticVLA (Ours) 0% 25% 50% 75% 100% 50% 55% 0% 0% 75% Jar pick-and-place 70% 85% 0% 0% 90% Waffles pick-and-place 50% 85% 0% 0% 95% Egg pick-and-place 56.7% 75.0% 0% 0% 86.7% Mean
n = 20 trials per task. An episode is successful if the object remains intact and is placed at the designated location.
02

Method

SA-RWFM + Tactile Distillation
HapticVLA method overview
Overview. Offline tactile reward computation feeds SA-RWFM teacher training alongside language, vision and state. At deployment, Tactile Distillation removes the tactile modality — the student predicts tactile representations from vision and proprioception alone.
Stage A

Offline tactile reward

For every episode we summarize left/right tactile maps into per-step safety rewards over force band, peak pressure, pressure concentration, inter-pad asymmetry, and slip detection, then combine with episode outcomes (success, drop, damage, risk).

Stage B

SA-RWFM teacher

A tactile-aware SmolVLA action expert is fine-tuned with reward-weighted flow matching — exponentiated clipped weights on a mixed episode / chunk advantage, robustly normalized per task, with an L2 anchor to the IL initialization.

Stage C

Tactile distillation

Teacher rollouts are precomputed once over the dataset. A tactile-free student is initialized from the teacher backbone (tactile encoder dropped, state projection sliced to proprioception) and trained on a 50/50 blend of ground-truth and teacher action chunks.

03

Hardware

Crab platform · bimanual SO-101
Crab robotics platform
Arms 2 × LeRobot SO-101 (6 DoF each). Left 7.4 V · right 12 V for added torque.
Gripper Custom parallel gripper, based on the open-source SO-ARM100/101 design, instrumented with tactile arrays.
Tactile 2 × 10×10 taxels → 200 taxels @ 120 Hz, 1–9 N per taxel.
Cameras Intel RealSense D435 (overhead) + 2 × IMX335 5 MP wrist. All streams 640 × 480 @ 15 FPS.
Compute Inference: Jetson Orin NX 16 GB. Training: RTX 4090 (SA-RWFM) · H100 (distillation).
Tasks Pick-and-place — marmalade jar · waffles · egg carton.
Tactile array
Tactile array · 100 taxels / finger
Example tactile map during manipulation
10×10 tactile map during grasp
04

Results

1-minute demo video
HapticVLA demo video Watch on YouTube →
Jar · Waffles · Egg — HapticVLA vs. baselines
05

Ablation — Tactile Distillation

n = 20 trials · per task

TD — Tactile Distillation. Each configuration combines the SA-RWFM teacher with or without distilling tactile representations into the student, and runs inference either synchronously or asynchronously.

Configuration Jar Waffles Egg Mean ↑
w/o TD · async 16 / 2018 / 2015 / 20 81.7%
w/o TD · sync 11 / 2017 / 2017 / 20 75.0%
w/ TD · async 14 / 2019 / 2015 / 20 80.0%
w/ TD · sync 15 / 2018 / 2019 / 20 86.7%

TD substantially improves contact-rich manipulation; combined with synchronous chunking it yields the largest mean success rate. Async inference underperforms its synchronous counterpart — we attribute this to temporal misalignment between tactile observations and control actions under async inference.

07

Citation

preprint · cite the arXiv version
@misc{hapticvla2026,
  title         = {HapticVLA: Contact-Rich Manipulation via Vision-Language-Action
                   Model without Inference-Time Tactile Sensing},
  author        = {Gubernatorov, Konstantin and Sannikov, Mikhail and
                   Mikhalchuk, Ilya and Fernando, Marcelino and Kuznetsov, Egor
                   and Ogunwoye, Faith Ouwatobi and Asanov, Artem and
                   Artemov, Makar and Guo, Ziang and Tsetserukou, Dzmitry},
  year          = {2026},
  eprint        = {2603.15257},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2603.15257}
}
08

Authors

Intelligent Space Robotics Laboratory · Skoltech
Konstantin Gubernatorov*Method · Platform · Experiments · Data
Mikhail Sannikov*Hardware · Software · Training · Distillation
Ilya Mikhalchuk*Hardware · Software · Inference · Data
Marcelino FernandoSimulation
Egor KuznetsovSA-RWFM
Ogunwoye Faith OuwatobiData · Experiments
Artem AsanovSimulation
Makar ArtemovDistillation
Ziang GuoAdvising
Dzmitry TsetserukouAdvising · PI
* Equal contribution · Intelligent Space Robotics Laboratory · Skolkovo Institute of Science and Technology, Moscow