Projects & Datasets

AgiBot World Colosseo is a full-stack and open-source embodied intelligence ecosystem. Based on our hardware platform AgiBot G1, we construct AgiBot World—an open-source robot manipulation dataset collected by more than 100 homogeneous robots, providing high-quality data for challenging tasks spanning a wide spectrum of real-life scenarios.

HoloPart is a novel diffusion-based model that completes partial 3D part segments into full, semantically meaningful parts, even when occluded. It combines local attention for fine-grained geometry and global context attention for shape consistency.

UniVLA is a unified vision-language-action framework that enables policy learning across different environments. By deriving task-centric latent actions in an unsupervised manner, UniVLA can leverage data from arbitrary embodiments and perspectives without action labels. After large-scale pretraining from videos, UniVLA develops a cross-embodiment generalist policy that can be readily deployed across various robots by learning an action decoding with minimal cost.

AnimateDiff enables animation generation from personalized text-to-image diffusion models by training plug-and-play motion modules. These modules learn transferable motion priors and use MotionLoRA for efficient adaptation without tuning the base models.

Vanilla autoregressive models without inductive biases on visual signals can achieve state-of-the-art image generation performance if scaling properly.

Janus is a novel autoregressive framework that unifies multimodal understanding and generation.

A scalable zero-shot 3D part segmentation framework that segments 3D objects into semantic parts at multiple granularities without requiring text prompts, enabling applications in robotics, 3D generation, and editing.

Scaffold-GS organizes sparse 3D Gaussians around structured anchors in space, dynamically predicting rendering attributes based on viewpoint and distance. It improves rendering quality and efficiency through structured anchor growth and pruning strategies.

PixArt-α is a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators.

SAM3D is a zero-shot framework that leverages SAM to segment 3D point clouds by projecting and merging 2D masks from posed RGB images, achieving fine-grained 3D segmentation without training.

T2I-CompBench is a comprehensive benchmark with 6,000 compositional prompts across attribute binding, object relationships, and complex compositions. It includes novel evaluation metrics and GORS, a fine-tuning approach to enhance compositional generation capabilities.

UniAD is a Unified Autonomous Driving algorithm framework following a planning-oriented philosophy. Instead of standalone modular design and multi-task learning, we cast a series of tasks, including perception, prediction and planning tasks hierarchically.

A paradigm for autonomous driving that applies both Transformer and Temporal structure to generate BEV features.

BungeeNeRF introduces a progressive training scheme that incrementally refines NeRF representations to support extreme multi-scale scene rendering, from city-scale context to high-detail objects.

Introducing AgiBot World, a large-scale platform comprising over 1 million trajectories across 217 tasks in five deployment scenarios. Accelerated by a standardized collection pipeline with human-in-the-loop verification, AgiBot World guarantees high-quality and diverse data distribution.

Facilitating the Perception, Prediction, Planning, Behavior, Motion tasks with human-written reasoning logic as a connection in between.

The largest driving video dataset to date, containing more than 1700 hours of real-world driving videos.