Researchers at ETH Zurich and collaborators developed a control approach that lets a humanoid robot climb ladders, cross obstacles, and recover from falls using a single learned policy.The system was demonstrated on Unitree’s G1 humanoid robot, showing how one approach can support different challenging movements. The method focuses on planning contact with the environment, rather than prescribing every movement in advance.In simulations, the approach achieved higher success rates than a comparison method across eight obstacle courses, highlighting its potential for improving humanoid robot mobility.Humanoid contact planningThe new system lets a humanoid robot climb ladders, cross obstacles, crawl through confined spaces, and recover from falls using a single learned policy.The technology, called Attention-based Contact Planning (ACP), helps the robot decide where to place its hands, feet, knees, elbows, and pelvis against the surrounding environment.The system was tested on Unitree’s G1 humanoid robot, demonstrating how contact planning and reinforcement learning can help robots adapt their movements to different terrain without following predefined motion sequences.With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier.Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data.Website: https://t.co/Y7cu3D4K4IPromo Video pic.twitter.com/ouZQskpZmQ— C. Zhang (@ChongZzZhang) October 23, 2026ACP combines reinforcement learning with motion imitation to address challenges in whole-body robot control. Reinforcement learning allows robots to improve through repeated attempts and feedback, while motion imitation teaches them to follow recorded movements. However, reinforcement learning can struggle with complex coordination, and imitation-based systems may struggle to adapt to unfamiliar obstacles.According to the team, instead of generating a complete sequence of body movements, ACP predicts suitable contact points on the surrounding terrain. A separate controller then coordinates the robot’s joints to reach those points.For example, when climbing a ladder, the robot must identify suitable rungs for its hands and feet. Its controller then coordinates its arms, legs, and torso to move between those points while maintaining support. The approach allows the robot to adapt its movements without following a fixed climbing sequence.The contact planner uses an attention-based neural network to process terrain information, the robot’s internal state, and a user-provided goal. Its inputs include joint positions, joint velocities, and body orientation. The system plans contacts for nine body parts: two hands, two elbows, two knees, two feet, and the pelvis. This allows the robot to use its upper body for support when walking alone is insufficient.Robots navigate obstaclesThe researchers also developed a LiDAR-based mapping system that converts sensor measurements into a local height map and estimates their reliability. This helps the robot identify narrow surfaces, such as ladder rungs, that might otherwise be missed. Running on a Jetson Orin NX computer, the system processes a 20,000-point LiDAR scan in 30 milliseconds, updating scans at 10 Hz.Researchers say the training took place in three stages. First, a whole-body controller learned to execute contact plans using 13.2 minutes of human motion data adapted to the robot’s body structure and physical limits. Next, the researchers trained the contact planner and controller together. Finally, reinforcement learning helped the system adapt to procedurally generated environments beyond the original demonstrations. Successful attempts provided additional training examples.During deployment, the robot does not need reference motions. It uses onboard observations and a specified goal to plan contacts and generate movements.Real-world demonstrations showed the G1 climbing ladders and stairs, crossing stepping stones, crawling, vaulting over obstacles, and recovering after falls. In simulations, the system achieved an average success rate of 92.5 percent across eight composite obstacle courses, compared with 58.8 percent for the MGMT baseline, a motion-generation-and-tracking approach. These figures represent simulated performance, not real-world success rates.The researchers note that the system currently focuses on terrain-based interactions and local movement toward a goal. Tasks involving overhead contacts, hanging, and manipulation would require more advanced three-dimensional perception.