TacInno-VLA

Official version

TacInno-VLA fine-tunes π₀.₅ for contact-rich robot manipulation. It uses images, instructions, robot state and touch to generate actions. A tactile tokenizer supplies contact-dynamics targets during training. The first release covers six UniVTAC policies and a four-task Franka FR3 policy.

View full model description

Papers & reports (1)

TacInno: Teaching Vision-Language-Action Models the Contact Dynamics Beyond Vision and Action

Paper available

arXiv: 2611.26405
Authors & affiliations (8)

Author order: Original order

  1. Weile Liequal_contribution

    ShanghaiTech University · Institute of AI for Industries (IAII), Chinese Academy of Sciences

  2. Yunfan Louequal_contribution

    Peking University

  3. Muqing Shiequal_contribution

    Washington University in St. Louis

  4. Zihao Zhang

    Institute of AI for Industries (IAII), Chinese Academy of Sciences

  5. Zhenbiao Huang

    Beijing FiveAges Technology

  6. Boxian Yao

    Peking University

  7. Xiyu Zhu

    Peking University

  8. Zhibo Pangcorresponding_author

    Peking University

First published
2026-11-20
Latest paper revision
v1 · 2026-11-20
Venue
arXiv

Covered model versions

Paper ↗PDF ↗

How to access and use

Public code resources
Public access
Model files
Public access

UniVTAC evaluation protocolUniVTAC simulation results (Table 1)Franka FR3 real-robot results (Table 4)

Code license

Code and training / fine-tuning

Access
Public access

SourceVerified on · 2026-11-19

Code license
Apache-2.0

SourceVerified on · 2026-11-19

Model files

Model files
Insert HoleAccessWeights license
Insert TubeAccessWeights license
Lift BottleAccessWeights license
Lift CanAccessWeights license
Pull-out KeyAccessWeights license
Put Bottle in ShelfAccessWeights license
Franka FR3 Four TasksAccessWeights license
Release information

SourceVerified on · 2026-11-19

Checks before deployment

The sensor’s exact Xense SKU is not specified in the real-robot documentation. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Inference mode
Closed-loop action chunks over WebSocket
Dataset format
HDF5 episodes converted to Zarr for simulation training
Runtime and frameworks
PyTorch
Fine-tuning requirements
π₀.₅ base weights, robot demonstrations and tactile code targets
Deployment requirements
Franka FR3, parallel gripper, one Xense vision-based tactile sensor, third-person and wrist cameras

Architecture and inputs / outputs

Model type
Action policy
Architecture
VLA
Backbone
π₀.₅
Action generation
Flow matching
Input
Language / RGB images / Robot state / Tactile
Output
Real robot: 7 joint target offsets and gripper opening
Released
November 19, 2026

Scope and training

Base model

π₀.₅

Code license

Apache-2.0

Implementation source

Official

Functional purpose

action_control

Record identity

Official version

Output

Actions

Prediction output

Real robot: 32 × 8 action chunk

Robot form

Robotic arm

Task scope

Multiple tasks

Training

Imitation learning / Supervised fine-tuning

Lineage

Fine-tuned from: π₀ · π₀.₅ ↗

Related robots

Selected checkpoint compatibility

Franka Research 3 ↗

Experimental · Other relationships

Wipe Board, Insert Plug, Lift Can, Unscrew Cap

Parallel gripper, one fingertip Xense vision-based tactile sensor, third-person camera and wrist camera

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Full parameters and verification notes (model card)

The encoder and decoder are conditioned on vision and executed actions. The tokenizer encodes the remaining future tactile change into four 16-value tokens; these supervise a code head alongside a contact-state head during training. At deployment the tokenizer and both auxiliary heads are removed. The action expert still reads 60 tactile tokens from three observations.

The project lists ShanghaiTech University, Institute of AI for Industries (IAII), Chinese Academy of Sciences, Peking University, Washington University in St. Louis, and Beijing FiveAges Technology as author affiliations. The arXiv preprint appeared in the November 20, 2026 batch.

Initial release timeline (Taipei time, November 19, 2026): initial code commit 19:27:30; weight repository 21:10:54; uploads 21:13:08–22:09:30; authors and affiliations updated 22:10:06. In Table 1, the authors report 87.5% across six UniVTAC tasks (100 episodes per task): hole insertion 95%, tube insertion 95%, bottle lift 78%, can lift 88%, key extraction 94%, and bottle shelving 75%. In Table 4, they report a Franka FR3 mean of 70.5% (50 trials per task): board wiping 88%, plug insertion 38%, standing a can upright 74%, and cap removal 82%. These are research results for the tested setup.

Seven downloadable policy checkpoints cover Insert Hole, Insert Tube, Lift Bottle, Lift Can, Pull-out Key, Put Bottle in Shelf, and the four Franka FR3 tasks. Six additional tokenizer packages contain two main checkpoints and four ablations. The Hub declares Gemma at repository level and explicitly applies base-model terms to the policy weights. Separate per-file license text for tokenizer weights and configuration files was not found. The public code documents simulation demonstrations on ModelScope; unauthenticated access could not be confirmed.

The evaluation guide gives the nominal seed range as 1000000–1000099 and specifies using the next seed when the simulator or planner errors.

Submit a source or correction