PolyUMI Video System VisTA Shape Recognition Object-in-Box Manipulation Code Contact
PolyUMI logo

PolyUMI

Accessible Visual–Tactile–Audio Data Collection for Object Inference and Manipulation

Conor W. Hayes*,1, Rickmer Krohn*,2,3,4, Aravind Ramaswami1, Anunth Ramaswami1,
Nils Dengler2,3,4, Kevin M. Lynch1, J. Edward Colgate1, Georgia Chalvatzaki2,3,4, Matthew L. Elwin1

*Equal contribution  ·  1Center for Robotics and Biosystems, Northwestern University
2Interactive Robot Perception & Learning (PEARL) Lab, TU Darmstadt  ·  3Hessian.AI  ·  4Robotics Institute Germany

The arXiv link goes live with the preprint.

Left: the three sensor streams — a fisheye wrist view of a lightbulb on a workbench, an iridescent tactile image of the contact surface, and an audio spectrogram with contact spikes. Right: the same sensing finger on the robot end-effector indoors and on the handheld gripper outdoors.
PolyUMI connects wireless demonstration collection to robot policy execution through a shared multimodal sensing finger. Wrist vision, optical tactile images and contact audio give complementary observations for learning contact-rich manipulation.
Abstract

Humans typically rely on vision, touch, hearing, and proprioception to perceive contact and adapt their actions during manipulation. Providing robots with comparable responsiveness therefore requires hardware that can retain and use these complementary sensory signals. Most imitation-learning systems, however, observe demonstrations primarily through vision and proprioception, limiting access to contact information that is difficult to infer visually.
We present PolyUMI, an open-source platform for scalable visual--tactile--audio demonstration collection and robot deployment. Its lightweight, wireless handheld gripper records synchronized wrist-camera, optical tactile, contact-audio, and proprioceptive observations without requiring a tethered workstation. The same sensing finger can be transferred to the robot end effector, preserving the sensing geometry between demonstration collection and policy execution.
To effectively use these heterogeneous observations, we further introduce VisTA, a token-level multimodal policy that integrates information across sensors and time to predict contact-aware robot actions. Experiments spanning object inference, slip control, and contact-rich manipulation show that touch and audio reveal task-relevant information beyond vision and that VisTA is competitive with or outperforms existing multimodal policies.
Together, PolyUMI and VisTA provide an accessible pipeline for collecting multimodal demonstrations and learning policies that perceive physical interaction beyond vision.

The system

A handheld gripper for collecting demonstrations and a robot end-effector for executing policies — built around one sensing finger that moves between them in about ten minutes.

Three panels. Left: the same scene seen by the wrist fisheye camera, the tactile camera and the contact microphone's spectrogram. Centre: the handheld gripper with labelled 177-degree fisheye camera, onboard compute and button-LED interface, conformant tactile surface and camera, piezo microphone and earbuds for live audio feedback. Right: the sensing finger mounted on a modified Franka Hand on a Franka Research 3 arm.
Left: the same scene through different sensors. The tactile camera sees the screwdriver through the sensing surface and the lightbulb through peripheral vision, while the wrist camera sees the bulb directly and the screwdriver is mostly occluded by the fingers. Contact events appear as spikes in the audio spectrogram and are heard live through headphones. Centre and right: the handheld gripper and the robot end-effector share the same tactile finger and task-oriented geometry.
The same sensing finger moving between the handheld gripper and the robot hand.
  • $236Sensing finger in parts, about $240 and four hours on top of a standard UMI gripper
  • 4Synchronized streams: wrist vision, optical tactile, contact audio, proprioception
  • ~10 minTo move the finger between the handheld device and the robot hand

Sensor modalities

The finger pairs an optical tactile sensor with a contact microphone in one module. An internal camera watches a deformable reflective surface through a curved mirror, giving a near-overhead view of a large contact region; the surface is seven layers of VHB tape coated in aluminium powder and sealed with medical tape. Dynamic lighting compensation keeps it usable across environments, which matters for in-the-wild collection.

ModalitySensorOutput
TactileOptical finger20 fps, 1152 × 648 MJPEG
AudioContact mic16 kHz mono PCM
VisionGoPro Hero 12 + fisheye60 fps, 1920 × 1080 MP4
ProprioceptionSLAM / ArUco6-DoF pose + gripper width

Cost and build time

The design uses modular components and minimal software dependencies, so the setup effort to deploy the system stays low.

ComponentHardwareAssembly
Robot end-effector$742.4830 min
Handheld UMI gripper$700.272 h
Multimodal sensing finger$235.964 h

Compared with other demonstration interfaces

PolyUMI is the only handheld interface that collects synchronized vision, touch and contact audio — and it adds live audio monitoring, so the demonstrator hears contact while recording.

MethodVisionTactileAudio Open sourceCable freeLive feedback
UMI✓✗✗✓✓✗
FastUMI✓✗✗✓✗✗
ViTaMin✓✓✗✗✓✗
Actuated UMI✓✓✗✓✓✗
ManiWAV✓✗✓✓✓✗
TacThru✓✓✗✓✗✗
Touch in the Wild✓✓✗✓✗✓
PolyUMI (ours)✓✓✓✓✓✓

VisTA

A token-level fusion policy. Instead of pooling one feature per sensor, VisTA keeps local tactile deformation, short auditory events and visual scene structure as separate tokens and lets attention mix them across sensors, space and time.

VisTA architecture: vision, tactile and audio streams each pass through a CNN encoder and the robot state through an MLP, producing modality-specific token sets that are fused into a single conditioning sequence for a flow-matching transformer that outputs an action chunk.
Sensor-specific CNN encoders produce 294 tokens in total, which a joint eight-layer transformer refines by bidirectional self-attention. The fused tokens C condition an 18-layer DiT-style action model trained with conditional flow matching.

Observations

Wrist camera and tactile images over a history of two timesteps, each resized to 224 × 224 and encoded to 98 tokens. Contact audio becomes a log-Mel spectrogram over roughly the last half second, encoded to 96 tokens. The robot state goes through an MLP. Modality-specific spatial embeddings and shared temporal embeddings keep the origin and time index of every token.

Actions

The policy predicts a chunk of 16 actions, each a 3D translation, a 6D rotation and a relative gripper-width command, expressed relative to the current end-effector pose. On the arm, 10 Hz commands are interpolated into joint torques at 1 kHz by a Cartesian impedance controller for compliant manipulation.

Experiments

Three questions. Do touch and audio expose object properties that vision misses? Do they give precise enough feedback for closed-loop slip control? And do they improve full manipulation behaviours?

  • Tactile

Tactile Shape Recognition

92.3% accuracy across five surface patterns

Fine surface geometry is hard to recover visually during manipulation: the contact region is occluded by the gripper and the features are smaller than an external camera can resolve. A CNN classifier trained on single tactile images separates five 3D-printed patterns, with scenes split between training and test so no scene appears in both.

Five iridescent tactile images showing the deformation signature of flat, square, circle, triangle and letter-N patterns.
What the finger sees: flat, square, circle, triangle and the letter N.
Confusion matrix for tactile shape classification, with diagonal values of 90, 99, 91, 84 and 97 percent.
Confusion matrix over the five classes.
  • Audio

Object-in-Box Classification

44% without audio → ~80% with it

The robot shakes a closed box containing gears, thin screws or thick screws and classifies the contents. Vision-only (44%), tactile-only (44%) and visual–tactile (42%) models land barely above the 33% chance rate. Adding contact audio lifts every sensor configuration to around 80%, with the tactile–audio models classifying thick screws perfectly — from only 20 training demonstrations per class.

The shaking motion: 20 oscillations along a sinusoidal trajectory, 10 cm amplitude, 0.15 s period.
Left: three audio spectrograms produced by shaking gears, thin screws and thick screws. Middle: photographs of the three object classes in the yellow box. Right: the box held by the robot gripper with a double-headed arrow showing the shaking motion.
The robot shakes the box through 20 oscillations. Each object class produces a distinct collision signature, visible in the spectrograms on the left.
Prediction distributions for object-in-box classification across seven sensor configurations. Configurations including audio reach around 80 percent accuracy; those without stay near 44 percent.
Prediction distributions per sensor configuration, aggregated over three random seeds. Green outlines mark correct predictions. V is wrist vision, T is optical tactile, A is contact audio.
  • Tactile
  • Audio

Closed-Loop Slip Control

8/10 with touch and audio, 2/10 with vision alone

The robot holds a screwdriver horizontally and must let it rotate to vertical by modulating grip force, without dropping it. Tighten too early and the screwdriver never reaches the target; react too late and it slips out. Doing this by teleoperation without haptic feedback is very hard, which is why the handheld platform matters for collecting the demonstrations at all.

A successful rollout: the policy lets the screwdriver rotate under gravity, then tightens to catch it near vertical.
SensorsSuccess rate
V + T + A8/10
T + A7/10
V + T3/10
V + A3/10
T3/10
V2/10
A0/10
Sensor ablation over 10 trials. Tactile and audio together enable slip control where other combinations fail.
Side view of the slip control task: the gripper holds a screwdriver which rotates from horizontal at the start, through the slip, to caught at vertical.
Start, slip, catch — the screwdriver must end within 10° of vertical, still in the gripper.
  • Vision
  • Tactile
  • Audio

Contact-Rich Manipulation

Two tasks with different demands. Board wiping needs sustained contact along a visually defined line: too little contact leaves the line unerased, deviation from the line gives only partial completion. Lightbulb turning chains alignment, insertion and rotational engagement — the robot must seat a loosely inserted bulb until it switches on.

Board wiping: a successful rollout, line fully erased.
Lightbulb turning: the bulb seats and switches on.
Grouped bar chart of success rates. On board wiping VisTA reaches 75 percent full success and 100 percent including partial, ahead of PolyTouch, Sparsh-X, MULSA and a vision-only diffusion policy. On lightbulb turning all methods exceed 80 percent.
Hatched regions are partial successes. On board wiping, the vision-only diffusion policy and MulSA frequently lose contact and Sparsh-X drifts off the line, while VisTA establishes contact, keeps it, and follows the line closely enough to erase it completely. On lightbulb turning every method clears 80% — the bulb state is a clear visual cue, so contact sensing adds little, and VisTA matches the strongest multimodal baseline without giving up performance.

BibTeX

@article{hayes2026polyumi,
  title   = {PolyUMI: Accessible Visual-Tactile-Audio Data Collection
             for Object Inference and Manipulation},
  author  = {Hayes, Conor W. and Krohn, Rickmer and Ramaswami, Aravind
             and Ramaswami, Anunth and Dengler, Nils and Lynch, Kevin M.
             and Colgate, J. Edward and Chalvatzaki, Georgia
             and Elwin, Matthew L.},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}

Update the arXiv identifier once the preprint is posted.

↑ Back to top