ByteDance Introduces Astra: A Dual-Model Architecture for Autonomous Robot Navigation

What happened

Traditional navigation systems typically consist of multiple, smaller, and often rule-based modules to handle the core challenges of target localization, self-localization, and path planning. Target localization involves understanding natural language or image cues to pinpoint a destination on a map. , QR codes). Path planning further divides into global planning for rough route generation and local planning for real-time obstacle avoidance and reaching intermediate waypoints.

The 4D spatio-temporal encoder replaces traditional mobile stack perception and prediction modules. It begins with a 3D spatial encoder that processes N omnidirectional images through a Vision Transformer (ViT) and Lift-Splat-Shoot to convert 2D image features into 3D voxel features. This 3D encoder is trained using self-supervised learning via 3D volumetric differentiable neural rendering.

The 4D spatio-temporal encoder then builds upon the 3D encoder, taking past voxel features and future timestamps as input to predict future voxel features through ResNet and DiT modules, providing current and future environmental representations for planning and odometry.

The planning head, based on pre-trained 4D features, robot speed, and task information, generates executable trajectories using Transformer-based flow matching. To prevent collisions, the planning head incorporates a masked ESDF loss (Euclidean Signed Distance Field). This loss calculates the ESDF of a 3D occupancy map and applies a 2D ground truth trajectory mask, significantly reducing collision rates.

Experiments demonstrate its superior performance in collision rate and overall score on out-of-distribution (OOD) datasets compared to other methods.

Extensive experiments were conducted in diverse indoor environments (warehouses, offices, homes) to comprehensively evaluate Astra’s performance.

Astra-Global’s multimodal localization capabilities were validated through various experiments, demonstrating superior performance in handling text and image localization queries. , “find the resting area”). Compared to traditional Visual Place Recognition (VPR) methods, Astra-Global exhibits significant advantages in: Astra-Local’s planning and odometry heads were thoroughly evaluated.

The planning head, using Transformer-based flow matching and masked ESDF loss, outperformed methods like ACT and diffusion policies in collision rate, speed, and overall score on OOD datasets. This highlights the masked ESDF loss’s effectiveness in mitigating collision risks.

The odometry head’s performance was assessed on multimodal datasets including synchronized image sequences, IMU, wheel data, and ground truth poses. Compared to two-frame BEV-ODOM baselines, Astra-Local’s odometry head showed significant advantages in multi-sensor fusion and pose estimation. Integrating IMU data dramatically improved rotational estimation accuracy, reducing overall trajectory error to approximately 2%. Further inclusion of wheel data enhanced scale stability and estimation accuracy, validating its superior multi-sensor data fusion capabilities.

For Astra-Local, improving robustness to out-of-distribution (OOD) scenarios is crucial, requiring enhanced model architectures and training methods. Redesigning the fallback system for tighter integration and seamless switching is also planned to improve system stability. Furthermore, integrating instruction-following capabilities will enable robots to understand and execute natural language commands, expanding their usability in dynamic, human-centric environments and fostering more natural human-robot interaction.

Why it matters

While foundation models have shown promise in integrating smaller models to tackle broader tasks, the optimal number of models and their effective integration for comprehensive navigation remained an open question.

ByteDance’s Astra, detailed in their paper “Astra: Toward General-Purpose Mobile Robots via Hierarchical Multimodal Learning” (website: https://astra-mobility.github.io/), addresses these limitations. Following the System 1/System 2 paradigm, Astra features two primary sub-models: Astra-Global and Astra-Local. Astra-Global handles low-frequency tasks like target and self-localization, while Astra-Local manages high-frequency tasks such as local path planning and odometry estimation. This architecture promises to revolutionize how robots navigate complex indoor spaces.

Astra-Global serves as the intelligent core of the Astra architecture, responsible for critical low-frequency tasks: self-localization and target localization. It functions as a Multimodal Large Language Model (MLLM), adept at processing both visual and linguistic inputs to achieve precise global positioning within a map. Its strength lies in utilizing a hybrid topological-semantic graph as contextual input, allowing the model to accurately locate positions based on query images or text prompts.

The construction of this robust localization system begins with offline mapping. The research team developed an offline method to build a hybrid topological-semantic graph G=(V,E,L): In practical localization, Astra-Global’s self-localization and target localization capabilities leverage a coarse-to-fine two-stage process for visual-language localization. The coarse stage analyzes input images and localization prompts, detects landmarks, establishes correspondence with a pre-built landmark map, and filters candidates based on visual consistency.

The fine stage then uses the query image and coarse output to sample reference map nodes from the offline map, comparing their visual and positional information to directly output the predicted pose.

For language-based target localization, the model interprets natural language instructions, identifies relevant landmarks using their functional descriptions within the map, and then leverages landmark-to-node association mechanisms to locate relevant nodes, retrieving target images and 6-DoF poses.

To empower Astra-Global with robust localization abilities, the team employed a meticulous training methodology. 5-VL as the backbone, they combined Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO). SFT involved diverse datasets for various tasks, including coarse and fine localization, co-visibility detection, and motion trend estimation.

In the GRPO phase, a rule-based reward function (including format, landmark extraction, map matching, and extra landmark rewards) was used to train for visual-language localization. 9% localization accuracy in unseen home environments, surpassing SFT-only methods.

Astra-Local acts as the intelligent assistant for Astra’s high-frequency tasks, a multi-task network capable of efficiently generating local paths and accurately estimating odometry from sensor data. Its architecture comprises three core components: a 4D spatio-temporal encoder, a planning head, and an odometry head.

What to watch

Astra holds significant promise for future development and applications. Its deployment can be expanded to more complex indoor environments like large shopping malls, hospitals, and libraries, where it can assist in tasks such as precise product location, efficient medical supply delivery, and book organization.

However, areas for improvement exist. For Astra-Global, while current map representations balance information loss and token length, they may occasionally lack critical semantic details. Future work will focus on alternative map compression methods to optimize efficiency while maximizing semantic information retention. Additionally, current single-frame localization can fail in feature-scarce or highly repetitive environments; future plans include active exploration mechanisms and temporal reasoning for more robust localization.