US robotics firm Figure has showcased its humanoid robots performing household tasks in a coordinated bedroom-cleaning demonstration. In a newly released video, two robots enter a minimalist room and begin organizing items, including hanging up a coat, closing a laptop, and placing headphones away.

The humanoids then work together to make a bed, coordinating their movements with head nods as they lift and smooth the comforter. The robots complete the task in under two minutes, highlighting advances in humanoid collaboration, object handling, and domestic automation.

Last month, Figure claimed to increase Figure 03 humanoid production from one robot daily to one hourly within four months at the new BotQ production facility in California.

Autonomous bedroom reset

Figure AI has demonstrated a major advance in collaborative humanoid robotics with two Helix-02-powered robots autonomously resetting a bedroom in under two minutes using a single learned Vision-Language-Action system. The robots performed a sequence of household tasks, including opening doors, hanging clothes, organizing objects, closing a book, disposing of trash, repositioning furniture, and jointly making a bed.

The demonstration focused on real-time coordination between two humanoids operating without a shared planner, central controller, or direct communication. Instead, each robot relied solely on its onboard cameras and learned policy to interpret the environment and infer the other robot’s intentions through movement. According to Figure, every action dynamically altered the scene, forcing both robots to continuously adapt their decisions while working toward a shared objective.

According to Figure, a key technical challenge involved manipulating a large deformable object — the comforter. Unlike rigid items, the bedding had no stable geometry or predefined grasp points. The robots had to predict each other’s actions while constantly adjusting grip, posture, and motion as the fabric folded, stretched, and shifted under shared tension.

The system also demonstrated advanced whole-body locomotion and dexterous manipulation. The humanoids balanced dynamically on one leg, used coordinated body posture to move furniture, operated foot pedals, and handled articulated and flexible objects without scripted transitions between tasks.

Vision-powered model

Figure claims its Helix AI framework, previously trained on logistics, laundry folding, and home-cleaning tasks, was scaled through additional data to enable collaborative multi-humanoid locomanipulation directly from visual input to physical action.

Figure recently unveiled a major upgrade to its Helix System AI model, adding perception-conditioned whole-body control to its humanoid robots. The update allows the robots to combine visual perception with body awareness, significantly improving navigation across complex environments such as stairs and uneven terrain.

Previously, the system relied only on proprioception, meaning the robots understood their movement and joint positions without visually interpreting their surroundings. With the latest version, S0 now processes input from onboard stereo cameras, transforming RGB imagery into a real-time 3D spatial understanding of the environment. This enables the humanoids to simultaneously “see” and “feel” the terrain while moving.

Figure said the system was trained end-to-end through reinforcement learning in simulation using highly randomized terrains and environmental conditions. The learned behaviors reportedly transferred directly from simulation to real-world operation without requiring additional calibration or fine-tuning, addressing one of robotics’ major sim-to-real challenges.

The upgraded control architecture enables the robots to walk with greater balance and stability across stairs and varied surfaces, even under changing lighting conditions. According to the company, the same perception-driven framework can support broader environment-aware behaviors beyond locomotion.