UniVTAC separates tactile simulation, representation learning, and policy evaluation

Release files and documentation checked September 19, 2026; paper analysis uses arXiv v1

UniVTAC is a February 2026 arXiv v1 preprint and public research project for simulation-based visuo-tactile data generation, tactile representation learning, and contact-rich manipulation evaluation. Its most useful contribution is not one headline score. It is a connected stack spanning a simulator, a synthetic encoder-pretraining corpus, a tactile representation encoder, an eight-task benchmark, public benchmark episodes, and a smaller physical sim-to-real study.

Those assets have different units and purposes. The **205,826 synthetic contact samples**, **400 paper-reported policy-training trajectories**, **800 HDF5 episodes in each current simulator-version task release**, and **450 physical demonstrations** are not interchangeable counts. The September file inventory also lists 638 contact-pretraining HDF5 files; it does not establish that their frames exactly reproduce the paper's pretraining corpus.

Start here: UniVTAC code and dataset downloads

Use the official project for demonstrations and the code repository for implementation. For data, select one of the pinned directories below before downloading. Inspecting the directory or our small CSV inventory needs no robot or NVIDIA GPU. Running the simulator has separate Linux and NVIDIA software requirements.

The inventory was generated from the provider's public file listing at revision e1aee7b0c95543b535e0146b2de3ee1bc6ddaabd. Its columns are checked_date, dataset_revision, component, group, hdf5_file_count, source_directory and verification_scope. Counts describe listed .hdf5 paths only: we did not download these HDF5 payloads, validate their arrays, total their frames or rerun policies. The CSV is RoboSkin's editorial inventory under CC BY 4.0; source data retain their own terms.

Choose the simulator version before installation

The project explicitly says the two task-data versions are not cross-compatible. Follow the pinned main installation guide or inspect the pinned 5.1 installer for your chosen branch. Both depend on the project's modified local TacEx source. Replacing it with an arbitrary latest TacEx or Isaac Lab release is not a verified installation path.

The 5.1 branch README still calls itself main in its installation paragraph; its branch notice and installer identify the 5.1 stack. We checked these files, but did not execute either installer or measure the README's throughput claim. The pinned download script uses ModelScope by default, while the inventory above audits Hugging Face. Its revision option is provider-specific: do not pass the Hugging Face revision to the ModelScope download command.

The paper lists ScaleLab at Shanghai Jiao Tong University, D-Robotics, ViTai Robotics, The University of Hong Kong, Nanjing University, Shenzhen University, Wuhan University, Fudan University, and Tsinghua University as its nine affiliations. These identify the authors' source-listed affiliations; they do not by themselves establish institutional ownership, endorsement, or responsibility for every project claim.

Short answer: what is UniVTAC?

UniVTAC has three named parts:

This distinction places UniVTAC in the middle of the Robot Skin → Tactile AI → Physical AI chain. A tactile sensor exposes local contact. The encoder turns tactile images into a model-ready representation. A task policy combines that representation with robot observations and actions. Physical trials then test whether the learned touch pathway changes behavior on a real robot.

Four data and evaluation pools that must remain separate

The paper and public repositories expose four materially different collections.

The simulation evaluation adds another unit: **100 rollouts per method-task pair**. The physical evaluation uses **20 rollouts per method-task pair**. Evaluation rollouts measure outcomes; they should not be added to training trajectories or described as more dataset episodes.

What the 205,826-sample corpus contains

The platform generates contact interactions with 14 geometric shapes. For each simulated contact sample, the source describes five supervision groups:

• tactile images with markers, `I_marked`;
• marker-free tactile images, `I_pure`;
• gelpad depth maps;
• projected marker coordinates; and
• a seven-dimensional relative object pose consisting of translation and quaternion orientation.

The paper reports randomized approach depth and small rotations so the corpus contains light-to-deep indentation, deformation, and shear-related marker displacement. These are simulation-privileged training signals. Their availability during encoder pretraining does not mean a physical sensor directly outputs every field during deployment.

That boundary matters for vision-based tactile sensing. Tactile images are observations; depth, pose, deformation, force-related cues, or slip risk can be simulator-provided or model-mediated quantities. They should not be presented as identical measurements.

The UniVTAC Encoder is a representation model, not a VLA

The shared encoder uses a ResNet-18 backbone and produces a 512-dimensional tactile feature. During pretraining, separate heads supervise three related capabilities:

1. **Shape:** reconstruct marked and marker-free tactile images.

2. **Contact:** reconstruct surface depth and projected marker positions.

3. **Pose:** regress the object's seven-dimensional pose relative to the gelpad frame.

The paper says the decoder heads are discarded for deployment and the retained encoder supplies tactile features to downstream policies. That makes UniVTAC Encoder a pretrained tactile representation component. It does not accept language and emit robot actions as a VLA model would, and it does not predict future world state as a robot world model would.

RoboSkin also does not classify it as a tactile foundation model merely because it is pretrained. A foundation-model claim needs explicit evidence about reuse and transfer across tasks, sensors, embodiments, or data domains. UniVTAC reports downstream tasks and one physical configuration, but the evidence remains tied to the paper's encoder, sensors, policies, and protocols. See tactile foundation models for the broader comparison standard.

Eight benchmark tasks and the paper protocol

The benchmark covers three source-organized capability groups:

In the paper's simulated policy experiment, each task has 50 full training trajectories. Each method-task pair is evaluated with 100 rollouts. The three rows in Table I use different tactile pathways: ACT without tactile input, VITaL, and ACT with the UniVTAC Encoder.

The change from 30.9% to 48.0% is **17.1 percentage points**. It is not a 17.1% relative gain, model accuracy, or a universal tactile-policy score. The eight task columns also vary substantially, which is why the average should not replace the task-level table in the primary source.

Physical sim-to-real evidence

The physical experiment uses a Tianji Marvin 7-DoF arm, a parallel gripper, a wrist RGB camera, and two ViTai GF225 tactile sensors sampled at 30 Hz. The paper describes Meta Quest-based teleoperation for collecting 150 demonstrations per task across Insert Tube, Insert USB, and Bottle Upright.

Each task and observation condition is trained separately. The authors then evaluate 20 rollouts per method-task pair under matching initial conditions and use human-observed task success.

These results are meaningful evidence that the synthetic-pretrained tactile representation changed success inside three physical task protocols. They do not establish a 25-point gain on other robots, sensors, objects, task distributions, or uncontrolled deployments. Human-observed success and 20-rollout samples should remain visible when the result is cited.

Public release audit: simulator, data, checkpoint, and licenses

The public release requires four separate checks.

The checkpoint logs inspected in the August 22 audit also do **not** reproduce the paper's Table I values exactly. Those UniVTAC checkpoint logs average 43.5, compared with 48.0 in the paper; the hosted vision baseline logs average 32.375, compared with 30.9 in the paper. The artifacts may still be useful, but these mismatches mean RoboSkin will not describe the public checkpoints as a reproduction of Table I without a reconciled protocol and result report. The September file inventory is not a rerun of these evaluations.

The public task data use the simulated GelSight Mini plus Franka Panda configuration. The physical study uses ViTai GF225 plus a Tianji Marvin arm. Mixing those hardware identities would hide the central sim-to-real boundary.

What UniVTAC contributes to tactile AI

UniVTAC connects three bottlenecks that are often discussed separately:

• scalable simulation data for tactile representation learning;
• a normalized task suite for testing whether touch changes policy behavior; and
• a physical experiment that probes transfer from synthetic tactile pretraining.

For robot learning, the project is a good example of why the unit of analysis matters. Encoder samples train a representation; trajectories train policies; rollouts evaluate policies; physical demonstrations adapt the system to a specific hardware setup. Counting all of them as “tactile data” without those roles would produce a larger but less meaningful number.

For robot skin, it shows how a fingertip tactile surface becomes useful only after sensor simulation or calibration, representation learning, synchronization with robot observations, policy integration, and task evaluation. For Physical AI, it demonstrates a contact-specific feedback path, not a complete general-purpose physical intelligence system.

What this paper and release do not establish

• The numerical paper analysis uses arXiv v1, submitted February 10, 2026. The official repository now announces CoRL 2026 acceptance; we have not independently checked a conference proceedings record or substituted a final paper's results.
• UniVTAC Encoder is not a VLA, a language-conditioned policy, or a demonstrated universal tactile foundation model.
• The 205,826-sample pretraining corpus is not the same counting unit as the 638 contact files or the 800 task files in each simulator-version directory.
• The 400 simulated policy-training trajectories are not the same as the 450 physical demonstrations.
• The physical 25-point average change is tied to three tasks, one research configuration, and 20 rollouts per method-task pair.
• The currently public simulator path does not establish complete working support for all three sensor types advertised by the project.
• The paper result and hosted checkpoint logs differ; an artifact download alone is not evidence of exact reproduction.
• Dataset, code, checkpoint, paper, and physical sensor software licensing must be audited independently.

FAQ

### Is UniVTAC a tactile foundation model?

RoboSkin classifies UniVTAC Encoder as a pretrained tactile representation encoder. The paper demonstrates reuse in its downstream policies and one sim-to-real setup, but it does not establish the breadth of cross-sensor, cross-robot, or open-task transfer needed to treat it as a general tactile foundation model.

### Is the public UniVTAC dataset 205,826 samples or 800 episodes?

The paper reports 205,826 contact samples for encoder pretraining. The September file inventory lists 800 HDF5 task files for each of two simulator versions, plus 638 contact files. File counts do not establish frame counts or equivalence to the paper corpus. None is the paper's 400 policy-training trajectories or 450 physical demonstrations.

### What is the UniVTAC Benchmark?

It is an eight-task simulated visuo-tactile manipulation suite covering pose reasoning, shape perception, and contact-rich interaction. The paper trains task policies with 50 trajectories per task and evaluates 100 rollouts per method-task pair.

### Does the public checkpoint reproduce the paper's 48.0 average?

Not in the hosted logs inspected in the August 22 audit. Those logs report a 43.5 UniVTAC average and 32.375 vision baseline, while the paper reports 48.0 and 30.9. The difference needs a protocol or artifact explanation before a reproduction claim is justified.

### Which robot and tactile sensor are used in the physical study?

The paper describes a Tianji Marvin 7-DoF arm with two ViTai GF225 tactile sensors and a wrist RGB camera. That physical configuration is different from the simulated Franka Panda and GelSight Mini configuration in the public task releases.

Related RoboSkin records

• Practice processing a small tactile dataset in Python
• Validate episode and timestamp fields before a LeRobot export
• UniVTAC pretraining corpus
• UniVTAC public benchmark data
• UniVTAC Benchmark
• UniVTAC Encoder model record
• ViTai GF225 sensor record
• Simulated Franka Panda and GelSight Mini configuration
• Physical Tianji Marvin and GF225 configuration
• Tactile AI
• Tactile manipulation
• Physical AI and touch

Primary sources

• Versioned arXiv HTML: UniVTAC v1
• arXiv abstract record
• arXiv-issued DOI
• Official UniVTAC project page
• Official UniVTAC GitHub repository
• Official UniVTAC Hugging Face dataset