A Common Navigation Architecture for Different Robots
Enabling robots to adapt to new environments and complete their routes safely remains among the primary technical challenges in autonomous technologies. NVIDIA introduced a new autonomous navigation framework called COMPASS to simplify this process and eliminate the need for retraining whenever the hardware changes. The architecture brings reinforcement learning and AI coding agents together while enabling machines with different physical characteristics to be trained on a common foundation.
High Success with Residual Reinforcement Learning
In conventional methods, whenever a new environment or robot design was introduced, all software interfaces, datasets and simulation environments had to be prepared again. NVIDIA COMPASS instead uses the pretrained X-Mobility navigation model. Through a residual reinforcement learning methodology, the system develops expert policies that add small corrective movements on top of this base model.
According to the test results, this approach increased the success rate by approximately 5 times compared with imitation learning models trained from scratch and significantly improved travel efficiency.
AI coding agents play important roles in automating the training and testing phases. These agents check software dependencies, create simulation assets, conduct initial tests and examine the failures that occur. Consistent with previous research showing that the key to success in AI agents lies in the control architecture, the COMPASS architecture prioritizes safety by retaining human approval at critical stages.
Direct Transition from Simulation to the Real World
The reference workflow makes use of various types of hardware, including the Boston Dynamics Spot quadruped robot, the Unitree G1 humanoid robot and the NVIDIA Carter wheeled robot. The NVIDIA SAGE-10K dataset, which contains 10,000 synthetic indoor environments covering 50 different room types, is used to prepare the test environments.
Omniverse NuRec technology is used to transfer real-world environments into the simulation environment and create precise digital twins.
Models trained only in simulation achieved an 80% success rate in complex environments when applied to real-world conditions without requiring additional fine-tuning. On robots using NVIDIA Jetson Orin as onboard hardware, inference time was measured at around 30 milliseconds. The optionally integrable NVIDIA cuVSLAM technology also provides visual odometry support to robots without external positioning data.