Keep pulling the thread on Tony Zhao and Cheng Chi.
Advances in AI perception allow modern robots to visually correct for their own mechanical inaccuracies, enabling the use of cheaper, less precise, and more compliant hardware components.
Sunday plans to launch a beta program in 2026, placing various prototype robots in people's homes to gather user feedback.
Sunday has deployed its data collection gloves to more than 500 people in real-world environments.
Sunday predicts that over one billion of its home robots will be in people's homes within the next few decades.
Sunday anticipates a potential commercial launch of its home robot in 2027 or 2028, contingent on the results of its 2026 beta program.
Sunday has collected nearly 10 million long-horizon trajectories of robotics data from real-world environments.
The Diffusion Policy algorithm enables robotics models to learn from multi-modal demonstrations, where a task is performed in multiple different ways, while maintaining training stability.
Sunday projects that the material cost for its home robot can be reduced to under $10,000 per unit at a production scale of a few thousand units.
Sunday's robot demonstrated zero-shot generalization by successfully performing tasks in six different, unseen Airbnb environments without any prior data from those locations.
The UMI project demonstrated that robotics data for imitation learning can be collected using a simple device like a GoPro attached to a 3D-printed gripper, without needing a full robot.
The current state of AI robotics is analogous to the period between the development of the core GPT technology and the launch of the consumer product ChatGPT, indicating a scalable recipe exists but has not yet been productized.
The performance of robotics models is expected to improve with increased data and model scale, following the same trajectory as other AI fields.