TARS

[TARS @TARSRobotics]

ICRA 2026 has wrapped in Vienna.

From DexHand’s global debut to AWE 3.0 live demos, TARS showcased a full-stack embodied AI system from data and models to dexterous manipulation.

Thank you to everyone who visited us. More to come.✨️

#TARS #ICRA #TARSDexHand #ICRA2026

[TARS @TARSRobotics]

The robotics world keeps asking: when does embodied AI actually work?

Today at #ICRA2026 Vienna, TARS Co-Founder and Chief Scientist Dr. Ding Wenchao answered it from the plenary stage.

AWE3.0: 3x higher success rate from novel viewpoints, 45% less jitter.

DexHand: 21 DoF, 0.05mm tactile resolution, 240Hz, full-hand control

Human-centric data: improving training utilisation.

Guinness World Record: sub-millimeter wire harness assembly in an actual industrial application environment.

His thesis: the near-zero-shot era of embodied AI is arriving earlier than expected.

Booth 163 · Hall B · June 1–5

TARS — the embodied brain that gets things done.🤖

#TARS #EmbodiedAI #Robotics #DingWenchao #AWE3 #DexHand #PhysicalAI #Vienna

[TARS @TARSRobotics]

TARS DexHand is more than a robotic hand.

Debuted at ICRA 2026, Vienna. It connects data, foundation model and quasi-direct-drive execution into a seamless hand-brain system.

From motion to manipulation — built to bring dexterous intelligence into the real world.

[TARS @TARSRobotics]

ICRA 2026 is underway, and the TARS booth is drawing strong attention.💥

From data and foundation models to dexterous manipulation, TARS is showcasing its full-stack embodied AI capabilities in Vienna.

Visit us at Hall B, Booth 163, Vienna Congress & Convention Center, June 1–5.

[TARS @TARSRobotics]

Just like a human mind needs a body, a world-class robot model needs an embodiment — hands included — designed to match it.

At ICRA 2026, TARS Robotics will present the 21-DoF TARS DexHand, our EAI-native dexterous hand built to connect embodied AI with the physical world.

Most robot hardware in Embodied AI 1.0 followed the same logic as industrial automation: The robot was designed first, and the AI adapted to it second.

That approach creates a fundamental mismatch.

When a model learns from human-centric data, the robot body must be able to reproduce human motion with precision and interact with the physical world through contact, force, and touch.

Our hardware is designed around the data distribution of human motion, the requirements of AWE 3.0, Visuo-Tactile World Model OmniVTA, and the physical demands of real manipulation tasks. The result is our EAI-native DexHand — one of the few robotic hands in the world to combine 21 degrees of freedom (DoF), quasi-direct drive actuation, and human-hand dimensions in a single integrated system. It is built for fine manipulation, tactile interaction, and contact-rich tasks that traditional grippers cannot complete.

Together with AWE 3.0 and OmniVTA, TARS DexHand doesn't just perform fine manipulation tasks but also understands and predicts contact.

This is the third pillar of Embodied AI 2.0: EAI-native embodiment, co-designed with the model and the data from the ground up. When all three work together, the gap between human capability and robotic capability starts to close.

Together with AWE 3.0 and SenseHub, TARS DexHand now completes the core loop of Embodied AI 2.0.

Hall B, Booth 163, Vienna Congress & Convention Center, June 1–5.

On June 4, our Co-Founder and Chief Scientist, Dr. Ding Wenchao, will deliver a keynote address on the Foundations for General Physical Intelligence.

#ICRA2026 #TARS #DexHand #EmbodiedAI2 #Robotics #EAINative #AWE #humancentric #sensehub #OmniVTA #mind2hand

[TARS @TARSRobotics]

Good data isn't enough. The architecture has to match it.

Embodied AI 1.0 often treated the robot brain as an extension of vision-language models: perception in, action out. This can work in controlled demos, but real environments introduce deformable objects, unexpected contact, changing viewpoints, and long-horizon task dependencies.

At ICRA 2026, TARS Robotics will showcase AWE 3.0, our general-purpose embodied AI model designed for the Embodied AI 2.0 era.

Rather than adapting a language model architecture, AWE 3.0 unifies vision, language, tactile sensing, and force feedback into a real-time world model.

Its key innovation: Before each movement, AWE 3.0 doesn't just map perception to action — it models what the physical world is about to do, considers multiple possibilities, and then decides. The result is a model that doesn't just execute specific tasks; it performs general tasks in real-world environments. This is what enables robots powered by AWE 3.0 to perform complex real-world tasks, including high-precision wire harness assembly, long-horizon packing, and dexterous manipulation.

In wire harness assembly — Goldbach's Conjecture of industrial automation, combining long-horizon reach, deformable object handling, and sub-millimeter precision — the robot powered by AWE 3.0 completed 100+ assemblies in a single hour.

A Guinness World Record.

This is the second pillar of Embodied AI 2.0: a model architecture native to the physical world, trained on Human-centric data, and built to get real tasks done.

Check it out for yourselves:

Hall B, Booth 163, Vienna Congress & Convention Center, June 1–5.

On June 4, our Co-Founder and Chief Scientist, Dr. Ding Wenchao, will deliver a keynote address on the Foundations for General Physical Intelligence.

#ICRA2026 #TARS #AWE #EmbodiedAI2 #humancentric #sensehub #Robotics #mind2hand

[TARS @TARSRobotics]

TARS Robotics is coming to ICRA 2026 in Vienna.

We are a leading embodied AI company, and the team behind the Guinness World Record for sub-millimeter wire harness assembly powered by AWE 3.0.

How did we do it?

The history of AI is the history of data...

Every breakthrough— from language models to vision-language models — has been unlocked first by better data. Embodied AI is no different. A robot doesn't just need to see the world. It needs to understand it, feel it, and learn to predict how the world responds to action. That demands data that captures force feedback, tactile sensations, contact states, and the full physical reality of human manipulation.

Instead of relying only on simulation or teleoperation, TARS took a different path: a human-centric paradigm that records real human actions as they happen, in real environments, with real physical feedback. The principle is simple.

Humans already perform millions of complex manipulation tasks every day.

Don't simulate that. Record it exactly as it happens, in real environments, with real physical feedback.

Our SenseHub data acquisition suite makes this possible. With a lightweight glove-and-camera system, SenseHub captures multimodal human manipulation data during natural work, including vision, motion, tactile signals, and interaction states, without disrupting the operator’s workflow.

This human-centric paradigm is why TARS was able to train AWE 3.0 — the world's leading general-purpose embodied AI model that actually works in real industrial environments— entirely on real-world human data.

We are convinced: 2026 is the critical year when human-centric data acquisition becomes the industry consensus.

Come and see SenseHub live at ICRA 2026 — Conference on Robotics and Automation in Vienna
— Hall B, Booth 163, Vienna Congress & Convention Center, June 1–5. #ICRA2026 #TARS #EmbodiedAI2 #humancentric #sensehub #Robotics

*On June 4, our Co-Founder and Chief Scientist, Dr. Ding Wenchao, will deliver a keynote address on the Foundations for General Physical Intelligence.