New AI system teaches humanoid robots to carry objects through cluttered spaces without stopping

What the system does#

LHM-Humanoid is a physics-based control system that enables a simulated humanoid robot to repeatedly pick up objects and transport them across cluttered environments in one continuous motion sequence. Rather than executing isolated, choreographed movements with resets between actions, the system orchestrates seamless cycles of walking, grasping, carrying around obstacles, and placing objects down—all while maintaining physical realism and balance.

The core challenge#

The hardest part isn't executing any single motion in isolation. Instead, it's managing the transitions between actions without interruption. When a humanoid places an object down, it ends up in an unstable, off-balance posture that bears little resemblance to canonical standing positions. From this awkward terminal state, the robot must both ensure it doesn't topple the object it just set down and somehow find a way to transition smoothly into the posture needed to walk toward the next item. Conventional end-to-end reinforcement learning breaks down in these scenarios because the learning algorithm has no easy path to success.

How it works#

The approach treats the transition problem as a two-sided recoverability challenge. The robot must learn to safely disengage from the object it just placed—preserving the success of that placement—while simultaneously settling into a balanced state from which the next fetch-carry-place cycle can begin.

The system uses two controllers working in tandem:

• First controller: Completes the full fetch-carry-place cycle and learns a "release-and-retreat" behavior that steers the final body state into a zone where recovery is possible.
• Second controller: Picks up from the resulting state distribution and launches the next cycle.

Both controllers are shaped by an adversarial motion prior that encourages realistic, human-like movement. The two policies are then distilled into a single unified goal-conditioned policy capable of running the entire sequence as one uninterrupted rollout.

Testing and results#

The researchers evaluated LHM-Humanoid across 350 cluttered room layouts spanning four different room types, comparing it against end-to-end reinforcement learning, hierarchical RL approaches, and prior physics-based human-scene-interaction methods. The system demonstrated significantly higher success rates and stability on both familiar and novel scenes.