Toyota Unveils CUE7 Demonstrating Advanced Basketball AI Vision

Toyota unveiled CUE7, its latest humanoid basketball robot, at Toyota Arena Tokyo, showcasing improved vision, balance, and shot consistency. The system demonstrates high-arc shots and dribbling routines, and follows a lineage that achieved a Guinness World Record free throw at 24.55 meters. The new hardware reduces mass from 120 kilograms to 74 kilograms and moves to an inverted two-wheel stance to support more dynamic movement.

CUE7 integrates high-fidelity sensing, trajectory optimization, and learned policies. The control stack combines reinforcement learning for skill acquisition with model predictive control for short-horizon stability and trajectory tracking. On the perception side the robot uses body-mounted sensors including lidar and vision to locate the hoop, estimate distance, and refine shot angle in real time. Hands and wrist actuation were redesigned for more consistent release mechanics and to reduce ball contact variability.

• •Hybrid control: reinforcement learning for policy robustness plus model predictive control for constrained actuation and balance
• •Mechanical changes: lighter frame, redesigned hands, single-wheel foot per side in an inverted two-wheel configuration
• •Sensing: torso lidar and vision stack for target localization and depth estimation
• •Training: sim-to-real loop where repeated simulator rollouts create diverse shot trajectories and recovery behaviors

Toyota is treating basketball as a compact, observable task that requires perception, planning, and coordinated whole-body control. That makes the court a convenient benchmark for iterating on locomotion, manipulation, and multi-sensor fusion without building a new task environment. The architecture reflects two converging trends in robotics: borrowing high-reliability perception and powertrain engineering from automotive R and D, and using end-to-end learned components where repetitive skill acquisition delivers robustness. CUE7 is not a research paper release, but it is a public demonstration of applied methods that bridge industry-grade sensing and modern learning approaches.

The hybrid use of learned policies for complex motor primitives combined with classical predictive control for safety and constraint satisfaction is a pragmatic pattern teams should note. Reducing mass and reworking mechanical degrees of freedom to simplify control while improving repeatability is a reminder that systems engineering remains central to fielded robotics. The visible improvements in shot consistency indicate practical sim-to-real strategies and data collection pipelines that scale beyond toy tasks.

Track follow-up technical writeups or repo releases from Toyota Frontier Research Center on their training pipeline, sim environments, and the exact sensor fusion approaches. Also watch for benchmarking of dynamic balance on uneven surfaces and for any transfer of these motion and perception subsystems into mobility or industrial applications.

"There is a widespread view that Japan is losing to China in physical AI, but we have created something that is not embarrassing to present to the world," said Tomohiro Nomi, research leader in Toyota's Humanoid Robot Research Group.