Two models, one robot
ByteDance has published a paper on Astra, a dual-model architecture for general-purpose mobile robot navigation in complex indoor environments. The system follows the System 1/System 2 cognitive paradigm, splitting navigation into two complementary subsystems: a slow, deliberative "global" model and a fast, reflexive "local" model.
Astra-Global handles low-frequency, high-complexity tasks: self-localization and interpreting user commands or images to identify navigation targets. Astra-Local handles high-frequency, rapid-response tasks: real-time local path planning and odometry estimation. The split mirrors how humans navigate, with deliberate route planning in the foreground and reflexive obstacle avoidance in the background.
How Astra-Global works
Astra-Global is a multimodal large language model built on Qwen2.5-VL that ingests images and text to perform both robot localization and target localization. It operates on a hybrid topological-semantic graph that encodes the spatial structure and semantic features of an environment using keyframes, landmark extraction, and node-edge relationships.
The training blends supervised fine-tuning with group-relative policy optimization, a reinforcement learning technique. The result is a system that achieves above 99% localization accuracy in new environments and outperforms traditional visual place recognition methods in robustness and detail sensitivity. In zero-shot generalization experiments, Astra achieved 99.9% localization accuracy in unseen home environments.
The practical benefit is that the system can disambiguate places that look identical to traditional methods, like two warehouse aisles that share the same shelving pattern, by attending to fine details like room numbers and specific landmark positions.
How Astra-Local works
Astra-Local is a multi-task network with a 4D spatio-temporal encoder at its core. The encoder processes series of omnidirectional images and sensor data to build a dynamic voxel-based environmental map for short-term planning.
The planning head uses flow matching, a generative modeling technique, and adds a masked ESDF loss to keep generated paths away from obstacles. The odometry head fuses IMU, wheel, and image data through a Transformer encoder to predict the robot's relative pose, with significant gains on rotation error and scale stability compared to traditional visual odometry.
What the benchmarks show
Compared to traditional Visual Place Recognition methods, Astra-Global demonstrated precise capture of fine details, stable localization even with large camera angle changes, and over 30% improvement in pose accuracy in warehouse environments. Astra-Local showed similar gains on planning and odometry.
Tests in simulated and real indoor environments, including warehouses, offices, and homes, demonstrated Astra's superior performance in localization, route planning, collision avoidance, and pose estimation compared to industry-standard approaches. ByteDance has deployed Astra on its own in-house mobile robots, and the system achieves a high end-to-end mission success rate across the test environments.
The paper positions Astra as a general-purpose architecture, not a benchmark-specific system. The authors argue that the System 1/System 2 split, combined with learned semantic understanding and 4D spatial reasoning, produces a robot that is robust to novel environments without requiring per-site retraining, which is the right shape of solution for a world where robots are increasingly expected to operate in human spaces rather than in carefully curated industrial settings.