ByteDance unveils Astra: A dual-module architecture for autonomous robot navigation

ByteDance has introduced Astra, a dual-model architecture for general-purpose mobile robot navigation, designed to overcome the bottlenecks of traditional navigation systems. Astra follows the System 1/System 2 paradigm with two sub-models: Astra-Global, a Multimodal Large Language Model (MLLM) handling low-frequency tasks like self-localization and target localization; and Astra-Local, a multi-task network for high-frequency tasks such as local path planning and odometry estimation. Astra-Global uses a hybrid topological-semantic graph and a coarse-to-fine two-stage process for visual-language localization, trained with Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) on Qwen2.5-VL backbone, achieving 99.9% localization accuracy in unseen home environments. Astra-Local comprises a 4D spatio-temporal encoder, a planning head with masked ESDF loss for collision avoidance, and an odometry head fusing IMU and wheel data via a Transformer model, reducing trajectory error to approximately 2%. Experiments in warehouses, offices, and homes demonstrated superior performance over traditional Visual Place Recognition (VPR) methods, with over 30% improvement in pose accuracy in warehouse settings. The team plans to extend deployment to more complex environments and improve robustness in out-of-distribution scenarios.