The Evolution of Generalist Embodied

Shanghai, China, March 10, 2025 (GLOBE NEWSWIRE) — Today, AgiBot launches Genie Operator-1 (GO-1), an innovative generalist embodied foundation model. GO-1 introduces the novel Vision-Language-Latent-Action (ViLLA) framework, combining a Vision-Language Model (VLM) and Mixture of Experts (MoE). The VLM utilizes internet-scale heterogeneous data to establish a solid foundation for scene and object understanding. The MoE consists of two key components: the Latent Planner, which learns from cross-embodiment and human operation data to develop general action understanding, and the Action Expert, which uses over a million real robot demonstrations to achieve high-frequency and dexterous manipulation. 

These components work in synergy, providing GO-1’s unique capabilities：

• Learning from Human Videos
• Few-shot Generalization
• Cross-Embodiment Adaptation
• Continuous Self-Evolution

Paper: https://agibot-world.com/blog/agibot_go1.pdf

YouTube Link: https://youtu.be/9dvygD4G93c

At the end of 2024, AgiBot launched the AgiBot World dataset, a large-scale, high-quality real world robotics dataset comprising over 1 million trajectories across 217 tasks in five application domains. Building on top of AgiBot World, today AgiBot introduces Genie Operator-1 (GO-1), a generalist embodied foundation model.

GO-1: An Evolution from VLA to ViLLA

To maximize the value of the high-quality AgiBot World dataset as well as web-scale heterogeneous videos while improving the policy’s generalization capability, AgiBot proposes a hierarchical Vision-Language-Latent-Action (ViLLA) framework. Compared to the Vision-Language-Action (VLA) model, where actions are directly conditioned on vision and language inputs, the ViLLA model predicts latent action tokens, bridging the gap between image-text inputs and robot actions generated by the action expert.

The ViLLA framework consists of a VLM and MoE. The VLM uses massive multimodal data on the internet to obtain general scene understanding and language comprehension. The Latent Planner in MoE harnesses data from various embodiments and human actions to build action comprehension. Meanwhile, the Action Expert, trained with over a million real world robot demonstrations, refines action execution. During inference, the VLM, Latent Planner, and Action Expert cooperate as follows:

1. VLM: Using the InternVL-2B model, it processes multi-view images, force signals and language inputs to provide scene understanding and instruction comprehension.
2. Latent Planner: This expert predicts Latent Action Tokens based on intermediate outputs from the VLM, forming a Chain of Planning (CoP) for general action understanding and planning.
3. Action Expert: It generates the final fine-grained action sequences based on intermediate outputs from the VLM and the Latent Action Tokens.

 

The following is an introduction to the two key components of MoE: Latent Planner and Action Expert.

 

Expert 1: Latent Planner

Although the AgiBot World dataset is the largest real world robot dataset globally, the volume of action-labeled robot data remains limited relative to internet-scale datasets. To address this, AgiBot employs latent actions to model the inverse dynamics of consecutive frames. This approach enables the transfer of real-world dynamics from heterogeneous data sources into universal manipulation knowledge.

• Latent Action Model (LAM): This model extracts the ground truth of Latent Actions between current and historical frames, consisting of an encoder and a decoder.

1. The encoder employs a spatial-temporal transformer with causal temporal masks.
2. The decoder uses a spatial transformer, taking the initial frame and discretizing Latent Action Tokens as input.
3. Latent Action Tokens are quantized using VQ-VAE.

• Latent Planner: The Latent Planner is responsible for predicting discrete Latent Action Tokens. It shares the same Transformer architecture as the VLM backbone but utilizes two independent sets of Feed-Forward Networks (FFN) and Q/K/V/O (Query, Key, Value, Output) projection matrices. The Latent Planner integrates intermediate VLM outputs layer-by-layer and  is trained using cross entropy loss.

Expert 2: Action Expert

To achieve high-frequency and dexterous manipulation, AgiBot integrates an action expert that utilizes a diffusion objective to model the continuous distribution of low-level actions.

• The Action Expert shares the same architectural design as the Latent Planner, utilizing the same Transformer backbone as the VLM but with two independent sets of Feed-Forward Networks (FFN) and Q/K/V/O (Query, Key, Value, Output) projection matrices. It employs a denoising process to iteratively regress the action sequence.
• The Action Expert is hierarchically integrated with the VLM and Latent Planner, ensuring consistency in information flow and collaborative optimization.

Experimental Results

 

Using the novel Vision-Language-Latent-Action (ViLLA) framework, AgiBot evaluated GO-1 across five tasks of varying complexity. Compared to current state-of-the-art models, GO-1 significantly outperforms them, increasing success rates by 32% (46% → 78%). Notably, tasks like “Pour Water” and “Restock Beverage” showed remarkable improvements. Furthermore, AgiBot validated the contribution of the Latent Planner within the ViLLA framework, showing a 12% success rate improvement (66% → 78%).

GO-1: Comprehensive Innovation of Embodied Intelligence

AgiBot GO-1 leverages human and diverse types of robot data, enabling robots to acquire revolutionary learning capabilities. It can generalize across various environments and objects, quickly adapt to new tasks, and learn new skills. At the same time, it can be deployed across various robotic embodiments, enabling efficient implementation and continuous evolution in real-world environments.

The key characteristics of GO-1 can be summarized as follows:

• Learning from Human Videos：GO-1 can learn from internet videos and real human demonstrations to enhance its understanding of human actions.
• Few-Shot Generalization：GO-1’s strong generalization ability enables fast adaptation to new scenes and tasks with minimal data, even in zero-shot scenarios, resulting in very low post-training costs.
• Cross-Embodiment Adaptation：GO-1 is a generalist robot policy model, capable of transferring between different kinds of robots and quickly adapting to various embodiments.
• Continuous Self-Evolution：GO-1 can continuously evolve from data generated by issues encountered during real-world execution, within AgiBot’s complete data feedback system.

The launch of GO-1 marks a rapid advancement of embodied intelligence towards generalization, openness, and enhanced capabilities: 

• From Single Task to Multi-Task: Robots can now perform multiple tasks across diverse scenarios without needing to retrain for each new task.
• From Closed Environments to Open Worlds: Robots are no longer limited to controlled lab settings but can operate in dynamic real-world environments.
• From Predefined Programs to Instruction Generalization: Robots can now understand and follow natural language instructions, reasoning and combining tasks based on semantics, rather than being confined to predefined programs.

AgiBot GO-1 will accelerate the widespread adoption of embodied intelligence, transforming robots from task-specific tools into autonomous agents with general intelligence. It will play a greater role across various domains, including manufacturing, service, and household applications, paving the way for a more versatile and intelligent future.

AgiBot official website：

https://www.linkedin.com/feed/update/urn:li:activity:7304747190139150338

https://fb.watch/yefx6B0bsC/