Atlas, the humanoid robot from Boston Dynamics, has recently showcased a significant advancement: it can utilize a single artificial intelligence (AI) model to master both walking and grasping tasks. This development marks a pivotal step towards creating general-purpose robotic algorithms.
The innovation arises from collaboration between Boston Dynamics and the Toyota Research Institute (TRI), which produced a generalist model that allows Atlas to control its arms and legs using a diverse set of examples. Traditionally, robots have needed separate models for different movements, but now Atlas can seamlessly integrate walking and grabbing, with its feet functioning almost like additional hands.
Russ Tedrake, a roboticist at TRI and MIT, noted that this integrated approach is impressive. Atlas’s learning model processes input from its visual and sensory data, as well as language prompts to execute various actions. It learns from multiple sources, including teleoperation, simulations, and demonstration videos, culminating in a large behavior model (LBM) that enables more natural performances. For instance, when reaching low to pick up an item, Atlas adjusts its legs for balance like a human would, and it even shows an emergent capability to recover dropped items independently.
This progress parallels the unexpected skills seen in large language models (LLMs), sparking hope among roboticists that similar techniques could result in a plethora of surprising abilities in robots. Tedrake’s team is also exploring different configurations of robotic arms for various tasks, hinting at a future where robots can efficiently learn new skills like cooking and cleaning without extensive retraining.
Although recent demonstrations of humanoid robots completing tasks appear impressive, such performances may involve extensive programming or teleoperation. Nevertheless, the advancements in Atlas signal a crossover in robotics akin to the development of generative AI models like ChatGPT. This could lead to robots capable of functioning in diverse conditions and learning rapidly.
While there’s still much to improve, experts like Ken Goldberg from UC Berkeley, who wasn’t directly involved in the Atlas project, acknowledge the significance of coordinating limb movements. However, he advises caution regarding the interpretation of emergent behaviors, as they may not be as novel as they seem.
Tedrake remains optimistic about impending breakthroughs in robotics, emphasizing the need for practical application of such technologies. As robotics approaches this pivotal moment, the potential for humanoid robots to engage in real-world tasks is increasingly tangible.
