Embodied AI: Teaching Robots to Understand the Physical World Through Experience
Current AI systems understand language and images but lack fundamental understanding of physics, objects, cause-and-effect, and the physical world. A robot trained only on internet data knows what a cup looks like but doesn't understand that cups hold liquid, that they break if dropped, that grabbing the handle is better than grabbing the rim. Embodied AI: systems that learn through interaction with the physical environment: is beginning to change this fundamental limitation.
In late 2025, a collaboration between Microsoft, MIT, and Boston Dynamics released PhysX-Learn, a framework for training robots through physical interaction. Robots learn concepts like fragility, friction, stability, and leverage by actually attempting tasks and receiving real-world feedback. A robot that's tried to grab 1,000 different objects develops intuitions about gripper force and angle that purely simulated training cannot teach.
Traditionally, robots were trained in simulation to avoid the expense and dangers of real-world training. The problem: simulation doesn't capture all the complexity of real physics. A friction coefficient that's correct in simulation is still different enough in reality that learned behaviors fail. This 'sim-to-real gap' is huge.
Research by OpenAI and others in 2024-2025 made progress on this through massive simulation plus careful real-world transfer. But embodied AI researchers like Yann LeCun argue the real solution is learning in the real world: accepting higher costs and failures as necessary for building systems that truly understand physics.
Physical robots learning through trial-and-error is slow and expensive. A robot might take weeks to learn a task that a human learns in minutes. But there's an emerging solution: learning from video. If you have video of humans performing tasks, AI can learn the patterns without physical experience. This cuts training time dramatically.
Boston Dynamics' humanoid robots, trained on massive video datasets combined with some physical experience, now perform industrial tasks with competence approaching human levels in specific constrained domains.
Please enable JavaScript to read the full article.