Telexistence Announces Results from Physical AI Fellowship 2026

Telexistence (TX) today announced the outcomes of its involvement in the 'Physical AI Fellowship 2026,' a collaborative initiative led by Amazon Web Services (AWS), NVIDIA, and MassRobotics. The company demonstrated a VLA-based autonomous operation by a humanoid robot and shared the results of its joint implementation of NVIDIA's World Model, 'DreamZero.' Centered on the theme 'From Real Humanoid Data to Physical AI,' the company showcased for the first time its end-to-end pipeline that directly links massive real-world teleoperation data to foundation model training to realize physical AI.

1. Single Policy Model Executes Autonomous Convenience Store Packing

The core of the demonstration is not a sequence of subtasks, but rather a single VLA model (single policy) that handles everything from perception to action end-to-end.

In the demo, TX’s humanoid robot retrieves multiple items (PET bottles, rice balls, snacks, etc.) at a Japanese convenience store checkout counter and packs them into bags. The demonstration marked the operations as 'Autonomous,' indicating they were performed without teleoperation intervention.

2. Joint Implementation with NVIDIA 'DreamZero': Bridging Real Data and Model Training

TX also revealed results from its joint initiative using NVIDIA’s World Model, 'DreamZero.' Through the Fellowship, TX worked with NVIDIA and AWS teams to implement DreamZero with a focus on practical application.

DreamZero is a World Model that predicts the 'future state' of a robot and its surrounding scene. In addition to motion prediction in simulated environments, the team conducted offline predictions from real-world humanoid teleoperation data, proving that data collected by TX in the real world can be effectively utilized for foundation model training.

By understanding real-world dynamics, World Models are expected to enhance Physical AI’s ability to respond to unknown events. TX continues this research with NVIDIA and AWS as a core focus of its robotics foundation model development.

3. Current Status and Future Challenges

TX has released this demonstration as a snapshot of its 'current research status' rather than a finished product. The current technical challenges are clear and constitute TX’s next development agenda:

- Operational Speed: The current autonomous operation is not yet at practical speeds. The demo showed footage at both 1x and 8x speed to frankly acknowledge that speed is the next major hurdle. - Motion Smoothness: Movements currently retain some awkwardness. The company is working on how to translate the VLA model’s output into final robot control, including action expression design. - Generalization: Precision drops when object positions shift by only a few centimeters. TX views this as a data volume and diversity issue, solvable through scale. Leveraging its unique position of handling everything from hardware design to real-world operations, TX is actively pursuing its own massive and diverse datasets.

4. Why join TX now?

TX is tackling these unprecedented challenges with several unique strengths: - Real Environment x Scale: Accumulating real-world data from operating robots in convenience stores, unavailable in academic labs. - Full Stack: Keeping hardware, data collection, foundation models, and control in-house allows for immediate reflection of developments in real machines. - Infrastructure: Conducting large-scale training using state-of-the-art GPU clusters from AWS and utilizing the IsaacSim environment in collaboration with NVIDIA.