Recently, Galaxea's open-source embodied foundation model G0.5 claimed the No. 1 spot on RoboDojo's latest Simulation Leaderboard, achieving an overall score of 20.23. The model delivered particularly strong results in two key categories, scoring 28.25 in Precision and 44.12 in Long-Horizon, demonstrating its capabilities in complex task understanding, continuous planning, and reliable execution.

Following its leading performances on internationally recognized benchmarks such as LIBERO and the BEHAVIOR-1K Challenge, G0.5's latest achievement on RoboDojo further validates the strengths of its unified autoregressive architecture in precision manipulation, long-horizon task execution, and cross-task generalization.

Topping a Public Benchmark: Validating General-Purpose Robotic Capabilities in Complex Tasks

Unlike traditional robotics benchmarks that primarily focus on success rates for individual tasks, RoboDojo places greater emphasis on a robot's comprehensive manipulation capabilities in realistic environments.

The platform provides standardized datasets, task suites, and evaluation protocols designed to systematically and objectively assess robotic models through 42 simulation tasks and 18 real-world robot tasks.

In this evaluation, G0.5 demonstrated particularly strong performance in Precision (28.25) and Long-Horizon (44.12).

In long-horizon planning and continuous manipulation, G0.5 maintained stable performance across tasks such as tabletop organization, tool storage, and multi-object manipulation. The robot could continuously interpret its environment, dynamically adjust its action strategies based on task progress, and ultimately complete the intended objectives.

G0.5 also demonstrated strong precision manipulation capabilities. When performing tasks involving insertion, alignment, and stacking, the model could not only identify target positions but also continuously refine motion trajectories based on environmental feedback, enabling stable and reliable execution.

Through large-scale, multi-task pretraining, G0.5 has developed transferable and compositional manipulation capabilities.

When encountering unfamiliar environments, new spatial layouts, or previously unseen objects, G0.5 can execute tasks directly from natural-language instructions, demonstrating strong zero-shot generalization and enabling robots to Think While Acting.

A Unified Autoregressive Architecture: Building a Foundation for General-Purpose Robotic Intelligence

G0.5's consistent success across public benchmarks is not the result of task-specific optimization. Instead, it reflects Galaxea's commitment to developing a unified autoregressive embodied foundation model.

Unlike conventional approaches that use a vision-language model as a separate encoder before passing information to a dedicated action module, G0.5 unifies visual understanding, embodied reasoning, and action generation within a shared Transformer decoder.

By processing task understanding and robot control through a single autoregressive sequence, the architecture allows reasoning to directly inform action decisions.

Building on this architecture, G0.5 further integrates embodied reasoning generation, structured action encoding, and spatiotemporal attention mechanisms.

As a result, robotic actions go beyond simply reacting to the current environment. The model can better understand task objectives, anticipate environmental changes, and continuously plan and execute actions.

This approach enables stronger task coherence during long-horizon operations while improving generalization across unfamiliar environments and tasks.

These capabilities have been validated across multiple public benchmarks.

To date, G0.5 has achieved leading results on LIBERO, BEHAVIOR-1K Challenge, RoboTwin 2.0, SimplerEnv-Bridge, Real-World Fine-Tuning, DROID Zero-shot Evaluation, and Pick-and-Place Benchmark, among others.

Its latest No. 1 ranking on the RoboDojo public leaderboard further demonstrates the potential of unified autoregressive architectures as a foundation for general-purpose robotic intelligence.

Meanwhile, Galaxea continues to advance its commitment to open-source development and collaboration in embodied AI.

G0.5's model, code, and related resources are now available to developers worldwide, lowering the barriers to embodied AI research and development and accelerating progress in robotic foundation models.

From public benchmarks to real-world applications, and from model capabilities to the broader developer ecosystem, Galaxea looks forward to working with researchers and developers around the world to explore the next frontier of embodied AI—enabling robots to truly understand the world and act autonomously within it.