ByteDance Introduces Astra — a Dual-Module Architecture for Autonomous Robot Navigation

ByteDance introduced Astra, a new architecture for mobile robot navigation that addresses the problems of traditional systems composed of many small rule-based modules. The architecture is based on the System 1/System 2 paradigm and includes two main models: Astra-Global, a multimodal large language model (MLLM) responsible for low-frequency tasks (self-localization and target localization), and Astra-Local, a multi-task network for high-frequency tasks (local path planning and odometry estimation). Astra-Global uses a hybrid topological-semantic graph built offline and a two-stage coarse-to-fine localization process. Training employed Supervised Fine-Tuning and Group Relative Policy Optimization. Astra-Local includes a 4D spatiotemporal encoder, a Transformer-based planning head with streaming alignment and masked ESDF loss, and an odometry head that fuses data from multiple sensors. Experiments showed that Astra achieves 99.9% localization accuracy in unfamiliar home environments, significantly outperforms traditional Visual Place Recognition methods in pose accuracy, and reduces collisions. The architecture is suitable for use in warehouses, offices, and homes, but improvements are needed for operation in low-feature environments and to increase robustness to out-of-distribution (OOD) scenarios.