As robots move deeper into warehouses, offices, and homes, one of the persistent obstacles has been getting them to reliably answer three basic questions: where am I, where am I going, and how do I get there. ByteDance has released Astra, a dual-model architecture designed specifically to address that gap for general-purpose mobile robots operating in complex indoor spaces.
Traditional navigation stacks typically string together several smaller, often rule-based modules to handle target localization, self-localization, and path planning separately. Self-localization is especially tricky in repetitive environments like warehouses, where systems have often leaned on artificial landmarks such as QR codes to know where they are. Astra takes a different approach, following a System 1 and System 2 style split into two sub-models: Astra-Global, which handles lower-frequency tasks like figuring out where the robot is and where a target destination sits on the map, and Astra-Local, which handles high-frequency tasks like real-time path planning and estimating the robot's own motion.
Astra-Global builds its understanding of an environment from a hybrid topological-semantic map constructed offline from video, where keyframes encode camera poses and are connected based on their relative positions, and semantic landmark information extracted from the visual data enriches each node. In practice, the model localizes itself or a target destination through a coarse-to-fine process, first identifying candidate landmarks and narrowing them down, then comparing fine visual and positional detail to pinpoint an exact pose. For natural-language target requests, such as finding a described location, Astra-Global links the instruction's meaning to landmarks stored in its map to retrieve the right destination.
On the local side, Astra-Local's design replaces traditional perception and prediction modules with a system built around omnidirectional camera images converted into 3D voxel representations, which are then projected forward in time to anticipate the environment a few moments ahead. Its planning component generates safe trajectories in real time while incorporating a distance-field-based loss designed specifically to cut down collision rates, and its odometry component fuses camera data with other sensors like IMU and wheel encoders to estimate the robot's own movement more precisely.
In testing across warehouses, offices, and homes, ByteDance reports Astra-Global reached 99.9 percent localization accuracy in previously unseen home environments after reinforcement learning was added on top of supervised fine-tuning, outperforming supervised-only training and traditional visual place recognition methods, particularly in scenes with repetitive layouts or large camera angle changes. Astra-Local's planning component reduced collision rates compared with alternative approaches on out-of-distribution test data, and its odometry component, once fused with additional sensor inputs, brought overall trajectory error down to around 2 percent.
ByteDance frames near-term applications around expanding into larger, more complex indoor spaces such as shopping malls, hospitals, and libraries, for tasks like locating specific products, delivering medical supplies, or organizing books. The team also flags open problems, including improving how the system handles feature-scarce or highly repetitive environments and building in stronger instruction-following so robots can act on more natural, conversational commands.