INTELLIGENT TECHNOLOGY

SOFTWARE

As humanoid robots move closer to everyday deployment, the focus has shifted from raw performance to safety.

RealSense has positioned its depth cameras and perception software as the‘ visual cortex’ for robots, enabling them to interpret surroundings, make decisions and move predictably in dynamic, humancentered spaces.

The demonstration, at NVIDIA GTC, was developed with LimX Dynamics and powered by NVIDIA technologies including visual odometry and cuVSLAM, showed how dense 3D perception allowed a legged robot to localise, map and navigate autonomously while maintaining stability and awareness.

Training and development were accelerated using NVIDIA Isaac Lab, a high-fidelity simulation environment that helped bridge the‘ sim-to-real’ gap. This approach enabled the humanoid system to learn complex locomotion and navigation behaviours in simulation before executing them safely in physical environments. localisation and mapping, allowing robots to understand their position and surroundings in real time.

It also highlighted reliable collision avoidance, ensuring safe interaction with people and moving objects, as well as stable locomotion across uneven terrain and elevation changes.

Unlike wheeled robots, which typically operate on flat, predictable surfaces, humanoids and quadrupeds must navigate full 3D space. Their movement involves shifting contact points, balance adjustments and non-linear motion patterns. Traditional approaches such as encoder-based odometry and 2D LiDAR lack the spatial awareness required for this level of complexity.

As a result, many legged robots have historically relied on teleoperation, human supervision or tightly controlled environments. These limitations have slowed broader adoption in real-world settings where unpredictability is the norm.

RealSense’ s approach combines dense depth sensing with advanced visual SLAM and NVIDIA’ s perception stack to deliver a more complete understanding of the environment.

This enabled the humanoid system to perform tasks that have traditionally been difficult to execute safely, including stair navigation, curb detection and traversal of uneven terrain.

The system also demonstrated dynamic obstacle avoidance, adapting in real time to moving objects such as carts, equipment and people entering its path. Advanced path planning allowed the robot to maintain smooth, humanreadable motion, reducing abrupt stops and erratic corrections.

The demonstration underscored RealSense’ s expanding role within the robotics ecosystem. Building on more than a decade of depth-sensing innovation, the company has developed technologies that support rapid prototyping, scalable deployment, and improved safety across robotic platforms.

As humanoid robotics continues to evolve, RealSense has positioned visionbased perception as a foundational layer for responsible autonomy. The GTC showcase illustrated how integrating perception, simulation and AI can accelerate the transition of humanoids from controlled demonstrations to practical, everyday applications. •

" Humanoids operate in constantly changing, three-dimensional environments," said RealSense CEO Nadav Orbach. " For robots to work safely alongside people, perception must go beyond sensing. It must enable understanding, stability, and predictable motion in unstructured settings."

The demonstration emphasised key requirements for safe robotic autonomy. These included continuous

28

INTELLIGENT CIO NORTH AMERICA www. intelligentcio. com