NVIDIA’s "DoorMan" Teaches Humanoids to Open Doors Faster Than Humans Can

Opening a door is a deceptive task. For a human, it is a subconscious act; for a robot, it is a nightmare of physics. The agent must navigate to the handle, identify it from a moving camera, grasp a rigid object, and then—crucially—coordinate its entire body to walk while manipulating a spring-loaded mechanism that pushes back.

In a new paper released this week, a team of researchers led by NVIDIA has demonstrated a system called DoorMan, which allows a humanoid robot to solve this "loco-manipulation" problem using only onboard cameras.

Remarkably, the system doesn't just work; it works better than we do. In real-world tests, the robot opened doors 23-31% faster than human expert teleoperators and achieved a higher success rate.

The research, which also features authors from UC Berkeley, CMU, and CUHK, was conducted on the Unitree G1, further cementing the $16,000 Chinese humanoid as the de facto development kit for the AI robotics era.

The "Pixel-to-Action" Breakthrough

Most current approaches to humanoid manipulation rely on "cheats" that are hard to scale: depth sensors that fail in bright sunlight, motion capture systems that require a studio, or massive amounts of human teleoperation data.

DoorMan takes a different path: Sim-to-Real Reinforcement Learning. The policy is trained entirely in NVIDIA’s Isaac Lab simulation environment and then deployed "zero-shot" to the physical world. The robot sees only RGB pixels—no depth maps, no markers.

To achieve this, the team solved two notoriously difficult problems in robotic learning:

• The Exploration Problem: In simulation, a robot learning from scratch will flail aimlessly for millions of attempts before it accidentally turns a handle. The researchers introduced a "staged-reset" mechanism. If the robot manages to grasp the handle, the simulator saves that state. In future training episodes, the robot is "reset" to that successful moment, allowing it to practice the opening and walking phases without having to re-solve the approach every time.
• The Visibility Problem: When a robot gets close enough to a door to open it, the handle often disappears from its camera’s field of view. The team used a technique called Group Relative Policy Optimization (GRPO) to fine-tune the student policy. This step forces the robot to learn behaviors that compensate for its own blind spots—like taking a step back or angling its head—to keep the task relevant features in view.

The Unitree G1: The Industry’s Standard Canvas

The choice of hardware for DoorMan is significant. The Unitree G1 has rapidly become the shared reality for robotics researchers, appearing in papers from Amazon’s Frontier AI & Robotics team and startups like Flexion.

Because the G1 is widely available and relatively standardized, it allows for direct comparison between methods. In the DoorMan study, the G1’s performance was pitted against human operators using VR controllers to drive the exact same robot.

The results were stark. Human teleoperators often struggled to "feel" the spring-loaded force of the door hinge through the VR interface, leading them to pull too hard or lose balance. The autonomous policy, having trained on millions of simulated doors with varying spring stiffness and damping, learned to be "compliant"—moving with the door rather than fighting it.

• DoorMan Success Rate: 83%
• Expert Human Teleop: 80%
• Non-Expert Human Teleop: 60%

Simulation as the "Multiverse"

The key to DoorMan’s success lies in the diversity of its training data. The team didn't just model one door; they procedurally generated a massive variety of them in Isaac Lab.

The system randomized everything: the width of the door, the damping of the hinge, the shape of the handle (lever, knob, push bar), and the texture of the materials. By exposing the AI to this "multiverse" of possibilities, the real world—with its specific lighting and physics—became just another variation within the model's distribution.

This aligns with the broader strategy NVIDIA outlined with its GR00T foundation models, which emphasize "physical AI" that learns from synthetic data rather than slow, expensive human demonstrations.