ACT-1: A Robot Foundation Model Trained on Zero Robot Data
ACT-1 successfully executed the 'Table-to-Dishwasher' task, involving 33 unique dexterous interactions and navigating over 130 feet. The model demonstrated zero-shot generalization by performing tasks in new homes without environment-specific training. ACT-1 advanced robot dexterity by completing tasks like folding socks and operating an espresso machine.
Evidence notes
- Sunday presents ACT-1 as a robot foundation model trained without teleoperation trajectories, using human-aligned Skill Capture Glove data and Skill Transform to scale manipulation learning.
- The release highlights ultra long-horizon mobile manipulation (e.g., table-to-dishwasher) and zero-shot generalization in unseen homes with map-conditioned navigation.
- The post is dated November 19, 2025 on the company journal and serves as the primary research narrative for the ACT-1 / Memo stack.
Company context
Sunday Robotics is building general-purpose robots designed to operate in home environments, combining hardware and control systems into a single platform. A core part of its approach is collecting real-world human data using glove-based systems, capturing how tasks are performed and using that data to train robot behaviour.