PsiBot, a robotics startup founded in 2024, has introduced its R1 model, an embodied AI system designed to perform complex physical and cognitive tasks in real-world settings. Unlike many current robots focused on basic motion functions, R1 is capable of playing Mahjong with humans, a game the company uses to showcase advanced manipulation and strategic reasoning.
Mahjong requires players to physically handle tiles while simultaneously planning moves based on incomplete information and opponents’ actions. PsiBot presents this as a Level 3 (L3) manipulation task, involving both dexterous action and autonomous, long-horizon decision-making. According to the company, L3 capabilities are essential for robots operating in dynamic, unstructured environments.
The R1 model is built on a hierarchical end-to-end architecture that separates planning and control functions but connects them through an internal mechanism called an “Action Tokenizer.” This design enables the system to interpret its surroundings, make decisions, and adapt its behavior during extended tasks. Reinforcement learning underpins the architecture, allowing the robot to improve performance based on prior experience.
In demonstration scenarios, R1 has maintained consistent reasoning and physical interaction for periods of up to 30 minutes while playing Mahjong, performing what PsiBot refers to as the “Chain of Action Thought” (CoAT) process. The company asserts that this loop—integrating perception, reasoning, and execution—marks a step toward the commercial deployment of general-purpose robots.
PsiBot states that the R1 platform is being tested for use in logistics, retail, and manufacturing environments. Partnerships with companies in these sectors are underway to explore how the technology can be applied outside the lab.
