Humanoid robot climbs ladders by itself

Brett Adcock, founder and CEO of Figure AI company in California, shared on X a video recording the scene of Figure 03 climbing an automatic ladder. Using the Helix AI model, this versatile humanoid robot can navigate unpredictable and ever-changing home environments.

Theo DPC Cars, climbing ladders is one of the most difficult tasks for humanoid robots because it requires dexterous coordination between arms and legs, constant balance adjustments, and precise perception of the surrounding environment. Each movement needs to be calculated to avoid losing balance when moving to the next step. Figure 03 successfully performed the task, partly demonstrating the robot’s ability to move, control the whole body and make real-time decisions.

Interesting Engineering reviews Figure 03 is getting closer to handling high-demand industrial tasks such as warehouse operations, equipment monitoring, and maintenance inside the factory. However, Figure AI has not described the technical details of the system that helps the robot climb ladders smoothly, and the video has not been independently verified.

Previously, this startup announced a major upgrade to the AI model Helix System 0 (S0), giving the humanoid robot the ability to combine visual perception with full-body motion control for better autonomous navigation. Meanwhile, the old system only relied on the robot’s perception of joint position, body movement and balance. Although effective for simple movements, that method limits the robot’s ability to safely navigate complex environments such as stairs, ladders, and less flat terrain.

After the update, the S0 now incorporates real-time vision data from the stereo cameras mounted on the robot. The color image processing model builds a three-dimensional simulation of the surrounding environment, allowing the robot to “see” the terrain ahead while continuously monitoring its own position. The combination helps the robot place its feet more accurately, improve balance and move smoothly on rough surfaces.

Figure AI shows that the model is trained from start to finish using reinforcement learning in an environment that simulates a series of terrains and random conditions. According to the company, learned behaviors transfer directly from the simulation to the actual robot without the need for further calibration or fine-tuning.

According to Brett Adcock, humanoid robots with legs are more suitable for environments associated with humans such as factories, warehouses, offices and homes. Figure AI is speeding up the production of Figure 03 from one robot per day to every hour. They delivered more than 350 robots, while improving production efficiency and developing AI.