Yunqi Capital | Charting the Path to World Models: Astribot Partners with Robotics Foundation Model Unicorn Physical Intelligence
How Far Is the "GPT Moment" for Embodied AI?
Enabling robots to "smoothly" understand and navigate the complex physical world is a major goal the embodied intelligence industry is working toward.
Recently, Astribot, a Yunqi Capital angel-round lead project, partnered with Physical Intelligence, a "unicorn" in general-purpose robotics foundation models, to collaborate at the model layer and jointly explore the path to world models. What new possibilities for embodied intelligence will this partnership unlock? This edition of Yunqi Partners walks you through the details.
01
Astribot S1 + π0
Generalization capabilities leveled up
On November 15, Astribot and Physical Intelligence jointly released a video demonstrating how the Astribot S1, enhanced with a collaboratively developed embodied model, smoothly operated a capsule coffee machine.
Whether retrieving coffee capsules, opening the machine's capsule compartment, or brewing coffee, the S1 completed tasks without any stuttering. Notably, in this demo, when instructed to make a Nopoli-flavored coffee, the S1 accurately grabbed the red Nopoli capsule regardless of where it was placed — demonstrating impressive generalization capabilities.
In fact, since launching its AI robot assistant Astribot S1 in August, Astribot has quickly gained attention for its outstanding all-around manipulation capabilities and embodied humanoid design. The Astribot S1 robot not only excels in multimodal perception, cognitive processing, real-time decision-making, and intelligent interactive execution, but has also made remarkable progress in AI training.
As one of the three core elements of AI, data is crucial for robot training. Astribot has distinctive advantages in this area, as it can leverage real-world data for model training at relatively low cost and high efficiency, collecting high-quality data across multiple dimensions — tactile, force, visual, and auditory feedback — from a first-person perspective, providing higher-dimensional data support for AI learning and execution.
Additionally, Astribot's "software-hardware integrated" architecture underpins its robots' industry-leading manipulation capabilities. From structural design to underlying hardware, the S1 pursues a balance between rigidity and softness, ensuring that in actual operation it can both control hard precision and grasp soft force, thereby accurately and flexibly adapting to operational demands across various scenarios.
Physical Intelligence, Astribot's partner in this collaboration, is also a rapidly rising startup in the robotics space, having announced a $400 million funding round in early November. The company focuses on developing general-purpose foundation models and learning algorithms, aiming to build a brain for robots. Its founding members include senior scientists from Google DeepMind, Stanford University professors, and UC Berkeley assistant professors. The π0 model released on October 31 is the first general-purpose robotics foundation model that Physical Intelligence developed over eight months.
The further evolution of the robotic model embedded in the S1 represents a joint effort by Astribot and Physical Intelligence to enhance and improve general-purpose robotics models.
Astribot's outstanding strengths in imitation learning and reinforcement learning, combined with Physical Intelligence's technical advances in multimodal embodied models, build a closed loop of embodied brain-cerebellum feedback iteration — this is the essence of their collaboration in jointly exploring world models.
02
Strong alliance
Exploring the path to world models
"When we founded the company in 2022, right when Yann LeCun had just proposed ideas around world models, we began thinking about robots and world models."
At the "BAAI Forum · 2024 Embodied Intelligence and World Models Summit" held in early November, Astribot founder and CEO Lai Jie stated that building world models is one of Astribot's goals. "Our robot collects and understands data between the world and the world model, then interacts with the physical world and feeds the results back to the world model."
World models are considered a critical path toward AGI and provide an important foundation for the generalization of embodied intelligence. By predicting future patterns, they enable deep understanding of both digital and physical worlds, allowing intelligent agents like robots to fully perceive, predict, plan, and adapt to complex and ever-changing scenarios.
World models remain in an early exploratory phase. As the lead investor in the angel round, Yunqi Capital welcomes and is optimistic about Astribot's joint R&D with innovative forces in general-purpose robotics models, and looks forward to new milestones in the construction of world models and the path toward robot AGI.
