In the past, when a robot was told, “Put the towel in the sink,” it could only execute a predefined sequence of actions. Now, with G0.5, the robot can reason through the task: Where is the towel? Where is the sink? It needs to pick up the towel, move toward the sink, and then place it inside. It thinks as it acts.
This capability is powered by our latest Vision-Language-Action (VLA) architecture. By enabling a single model with a single set of weights to handle both reasoning and action generation, G0.5 gives robots the ability to think while acting.
1. From Language to Action: Zero-Shot Generalization to New Objects, Environments, and Instructions
On the R1 Lite robot, G0.5 can perform manipulation tasks directly from natural-language instructions, without any fine-tuning for the target scene. Even when the model has never encountered the current environment layout, object categories, or instruction combination, it can interpret the task, decompose it into subtasks, and generate smooth bimanual control in real time.
This is G0.5’s ready-to-use capability in action: one model, one instruction, and the robot gets to work.
2. From Memorizing Tasks to Learning How to Manipulate
Traditionally, teaching a robot a new task often means collecting new data, fine-tuning the model, and adapting it to the target environment. Robots may perform well on tasks they have been trained on, yet struggle to generalize to unfamiliar objects, environments, and instructions.
Through large-scale, multitask pretraining, we have enabled G0.5 to develop transferable and composable manipulation capabilities. Fundamental skills such as grasping, placing, pushing, pulling, opening, closing, and moving are no longer fixed steps tied to individual tasks. Instead, they become reusable building blocks that the model can combine to handle new situations.
When faced with unfamiliar scenes, spatial layouts, and previously unseen objects, G0.5 can act directly on natural-language instructions. Robots begin to Think While Acting.
3. Three Core Capabilities That Enable Robots to Think While Acting
3.1 A Unified Heterogeneous Action Tokenizer: One Action Language for Different Robot Embodiments
Robots vary widely in morphology, degrees of freedom, and control frequency. Bimanual humanoids, mobile manipulation platforms, and single-arm industrial robots all have different action dimensions.
Existing approaches often take one of two paths. One is to develop a separate pipeline for each embodiment—for example, using a fixed DCT-based pipeline in FAST to discretize actions independently for different embodiments. The other is to flatten all degrees of freedom into a single long vector before quantization.
Both approaches have limitations. The first makes it difficult to share knowledge across embodiments. The second causes the number of tokens to grow linearly with the total number of degrees of freedom. Even when only a finger needs to move, the model may still need to generate tokens for every joint. This creates unnecessary overhead and leads to entangled action representations that are harder to transfer across tasks and embodiments.
G0.5 Action tokenizer
We believe action tokens should be both structured and sparse, much like language.
To achieve this, we developed a unified heterogeneous action tokenizer. With a shared action vocabulary, G0.5 can represent a wide range of embodiments, from tabletop bimanual manipulation to whole-body mobile manipulation. Meanwhile, the additional overhead of autoregressive decoding remains manageable.
This design is key to putting the vision-language model (VLM) back in the role of an action generator, rather than reducing it to a conditional encoder.
3.2 Native Action Chain-of-Thought: Enabling Robots to Reason About How to Act
A unified action vocabulary allows us to put the VLM back at the center of action generation. The real benefit is that the generative capabilities learned during VLM pretraining—including chain-of-thought reasoning, in-context learning, and prompt-based control—can be applied directly to action generation, without passing through the information bottleneck of the VLM-as-Encoder architecture.
In many existing approaches, chain-of-thought (CoT) is used only as an auxiliary training objective. Reasoning and action generation are separated into different parameter sets and objectives. For reasoning to influence action, its output must first be compressed into a latent representation and passed to a downstream action expert. This introduces potential semantic information loss and makes latency less predictable.
We take a different approach: integrating CoT and action generation into a single autoregressive stream. The model first generates reasoning outputs—including subtask decomposition, target-object bounding boxes, and 2D trajectory hints—and then generates action tokens.
This design delivers two distinct, measurable benefits.
- Zero-Shot Decomposition of Long-Horizon Tasks
Household tasks in benchmarks such as BEHAVIOR-1K require natural-language instructions to be decomposed into dozens of subtasks. With just one epoch of training, a single G0.5 checkpoint outperforms π0.5 trained for four epochs, as well as the challenge-winning solution composed of four checkpoints. G0.5 can also decompose subtasks zero-shot in household tasks outside its pretraining distribution.
- Real-time behavioral control through language
Because prompts enter the same autoregressive stream as action tokens, natural language can continuously shape the model's action distribution at inference time, without retraining.
Consider the task of turning on a toaster. Because the switch requires a relatively long press, the model may not apply enough downward force when instructed only to “turn on the switch.” Adding “push harder” to the instruction leads the model to increase the downward force and successfully activate the switch.
G0.5 performs the task of placing a towel in a sink zero-shot on R1 Lite. Within the same autoregressive stream, the model first generates reasoning outputs, including subtasks and target-object bounding boxes, before generating action tokens. It continuously replans based on observations from each frame.
3.3 Spatiotemporal Attention: Giving Robots Context-Aware Perception
Complex household tasks cannot be solved reliably through reflexive responses to a single frame. When a robot's view is briefly obstructed by its own arm, or an unexpected failure occurs, it needs historical context to maintain a consistent understanding of the scene.
G0.5 incorporates a lightweight spatiotemporal attention module that fuses several seconds of visual history. This enables the model to operate more robustly even when parts of its immediate visual field are temporarily unavailable.
Experiments show that, supported by perceptual priors learned during pretraining, G0.5 consistently outperforms π0.5 on long-horizon BEHAVIOR-1K tasks such as moving a box to a storage room, loading a vehicle, carrying firewood, and tidying a bedroom.
4. Strong Results Across Seven Evaluation Settings
G0.5 demonstrates leading performance across seven independent evaluation settings, with key results summarized below:
Real-world fine-tuning evaluations: Across six task–embodiment configurations on the R1 Pro and R1 Lite platforms, G0.5 (dark orange) delivers substantially stronger overall results than π0.5 and GR00T-N1.7 in both task success rates and process scores.
DROID zero-shot evaluation: Without any fine-tuning on the target robot, G0.5 achieves an average success rate of 82.5% across 10 tabletop manipulation tasks. It outperforms both π0.5-DROID and MolmoAct2-DROID on every evaluated task.
Pick-and-place benchmark: From zero-shot evaluation through 50 hours of post-training, G0.5 (dark-colored bars) outperforms π0.5 across all evaluated data scales on both language-following rate and task success rate.
Taken together, these results highlight several important capabilities.
-
Large-scale pretraining builds core embodied foundation model capabilities. With lightweight post-training, G0.5 surpasses π0.5 across seven benchmarks and achieves state-of-the-art (SOTA) results in the reported evaluations, validating the effectiveness of this pretraining approach.
-
Zero-shot, ready-to-use generalization: Zero-shot experiments on DROID (Franka) and PP Bench (R1 Lite) show that G0.5’s pretrained capabilities can transfer directly to compatible robot platforms and a previously unseen environment.
-
Single-model generalization and performance advantages: Across the 50 long-horizon mobile manipulation tasks in the BEHAVIOR-1K Challenge, G0.5 uses a single set of model weights to achieve a score of 0.29 after just one epoch of post-training, significantly outperforming both the challenge-winning multi-model ensemble and π0.5. After four epochs, its score rises to 0.31, demonstrating further performance gains, with G0.5 outperforming baselines on more than half of the tasks.
-
Pretrained representations overcome long-horizon task bottlenecks: These results directly demonstrate that pretrained priors built on structured action spaces and visual memory are central to bridging the gap in long-horizon mobile manipulation and enabling efficient generalization.
5. Conclusion: From Executing Actions to Understanding the World
G0.5 marks an important step forward in Galaxea’s embodied foundation model development.
Rather than using the VLM solely as an encoder, G0.5 puts it back at the center of action generation.
Rather than separating reasoning from action, it enables the model to think while acting.
Rather than simply executing predefined routines, it enables robots to understand how to act and retain context about what has happened.
We believe that general-purpose embodied intelligence requires a scalable path for both models and data. We will continue advancing this approach with more robot data, more complex environments, and longer task horizons.
When the model weights are released, G0.5 is intended to support ready-to-use deployment across multiple robot embodiments, helping developers explore practical applications of embodied AI.
Explore the G0.5 technical report:
0 条评论