Bringing a Frontier World Model to the Convenience Store: Inside Telexistence’s DreamZero Experiment on AWS

This article describes how Telexistence, a Japanese robotics startup, collaborated with AWS to evaluate NVIDIA's DreamZero world action model for retail robot manipulation tasks, building a complete cloud-native pipeline on AWS infrastructure.

• DreamZero is a 14-billion-parameter world action model that predicts future camera views alongside actions, enabling robots to anticipate consequences before moving.
• Built an end-to-end pipeline on AWS spanning data conversion, GenAI-based curation, distributed fine-tuning on B200 GPUs, and multi-pronged evaluation.
• Developed a multi-agent GenAI curation pipeline that automatically filters noisy teleoperation data, flagging low-quality demonstrations with 39.3% accuracy versus 49.7% human review.
• Successfully fine-tuned DreamZero across embodiments from AgiBot G1 checkpoint to Telexistence's TX-G2 robot, achieving 45% closed-loop success on clean simulation data.
• Identified real-to-sim gap challenges and optimized inference latency to ~583ms per step on B200 instances, making real-time control feasible.

The PoC demonstrates how rigorous cloud-scale evaluation provides evidence-based insights for adopting frontier robotics models, with reusable pipeline components applicable to other robot-foundation-model builders.