What Does a Post-AGI Physical World Look Like?
Ted, founder of Prometheus, discusses the verification bottlenecks between AGI and physical abundance at a Humanoid Hub event hosted by Devang.
Xiaoen, CEO and founder, and Suen, chief scientist, of Delta Intelligence discuss the Delta-0 humanoid foundation model with RoboPapers hosts Michael and Chris.
Suen presents demonstrations using a Unitree G1 with U2 hands, including playing a record, making a bed, picking objects up, and opening a dishwasher. One sequence runs about three minutes, with nine tasks and more than 30 steps. The team says the robot handles these tasks through one model, without SLAM-based navigation.
Delta Intelligence treats tasks that combine movement and object handling as a central robotics challenge. Actions change the robot’s position and surroundings, so each step must set up the next while accounting for different viewpoints, body poses, contacts, and object conditions. For example, the G1 braces against a surface to get enough force to open a dishwasher.
Delta-0 pairs a vision-and-instruction model with a whole-body controller trained through reinforcement learning in simulation. The model predicts motion commands, and the controller turns them into robot control signals. Delta Action Control lets a person adjust a rollout with small changes to the robot’s position, which the team says can support human-guided learning.
The team reports training with about 10,000 hours of human and egocentric data, and aims to grow that to 100,000 hours or more. Its collection setup adds a downward-facing fisheye camera to capture more of the person’s body, including movements missed by typical first-person recordings. Delta Intelligence also builds simulated room replicas to compare robot performance in simulation and in the real world.
Delta-0 uses human interventions during robot rollouts to gather experience for retraining policies and value models. In one out-of-distribution task, the team reports success rising from about 20% to around 65% after using intervention data. Xiaoen says the company is focused on improving reliability, speed, and deployment, and is hiring people interested in humanoid robotics.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Ted, founder of Prometheus, discusses the verification bottlenecks between AGI and physical abundance at a Humanoid Hub event hosted by Devang.
Do Park of Lightberry explains how the company designs Lumi for emotionally engaging human interaction at The Humanoid Hub.
Aadeel Akhtar of PSYONIC presents at Humanoids Summit on bionic hands that serve prosthetic users and help robots learn manipulation tasks.
Light Origins’ presenter explains how Light-O1 uses human video to train robot action models and demonstrates it on a Lightwheel setup and a Unitree G1.
Joseph Amigo, a second-year PhD student, and Ruhollah, a fourth-year PhD student at NYU, discuss their paper on robot reinforcement learning with RoboPapers hosts.
Armin Foshnaughski of Munro & Associates, Schaeffler’s Al Mackie and Paul, and Scott Walter of RoboStrategy discuss the Unitree G1’s design and humanoid robotics during a Munro teardown.