Google DeepMind unveils method to teach new skills to Apptronik’s Apollo from video

Robotics & Automation News reported on October 22, 2026, that Google DeepMind had demonstrated a method for teaching new skills from video to Apptronik’s humanoid robot Apollo using Gemini 4 Argon. According to the article, the video showed a game of cornhole. The model analyzed the environment, assigned coordinates, calculated full-body joint movements and autonomously generated code for a simulation. The learned behavior was then transferred to Apollo, which reportedly made seven consecutive successful throws in real-world trials.

Google DeepMind says the method can cut the time needed to create a simulation from several weeks of manual work to a few hours. If its reproducibility is confirmed as the company expects, the approach could shorten the robot training process itself.

The focus is not simply that the robot was remotely operated, but that a single workflow linked video to perception, spatial estimation, motion planning, simulation code generation and transfer to a physical robot. Apptronik’s Apollo features in the demonstration, but the main question is less the robot’s own performance than how much of the robot-learning process a general-purpose AI model can take on. Google DeepMind has also showcased robotics demonstrations in Google demonstrates Apollo robotics trials with Gemini 4 Argon. This latest demonstration appears to extend that work by showing a specific path from video to a physical robot. Related coverage by this newspaper underscores how the news sits alongside debates over the safety of robot planning and video-based control. A report on research from Binghamton University, Research examines safety of AI-powered robots, focused on the risks of incorporating LLMs into robot planning. The latest method combines video understanding with code generation to shorten training, but it relies on Google DeepMind’s internal trials, and the article does not explain how reproducibility or safety will be verified. See the original report for details. At the industry level, the focus is shifting from the manufacturing capabilities of individual robots toward software platforms that link video input, simulation generation and transfer to physical robots. If this approach moves closer to practical deployment, robot development could shift from manual design for individual tasks toward processes centered on data and generated code. However, the demonstration described here covers only a single task, cornhole; whether it can be extended directly to general warehouse or factory work is a separate question. The next points to examine are which tasks can be reproduced, what conditions apply to the training videos and how errors in simulation generation can be controlled.

For Japan’s robotics industry, competitive advantage may increasingly depend not only on the mechanical design of individual robots but also on how processes link video data, simulation and transfer to physical robots. Based on the company’s description, businesses with platforms that can generate simulations from video may be able to accelerate prototyping.