Awesome Galaxea Developers

The OpenGalaxea developer ecosystem is continuously updated with new community projects and contributions. Check back often to explore the latest tools, demos, and research powered by OpenGalaxea.

🤖 Research & Papers

Project Name	Tags	Description
BEHAVIOR Robot Suite	#whole body control #mobile manipulation	Developed by Stanford in collaboration with Galaxea, BRS leverages the Galaxea R1 robot to enable whole-body mobile manipulation for diverse household tasks. It has been accepted by CoRL 2025.
DenseMatcher	#manipulation	Developed with the Galaxea A1 robotic arm, DenseMatcher enables cross-instance 3D semantic understanding, allowing robots to learn tasks like peeling a banana from a single human demonstration. This paper has been accepted by ICLR 2025 as a Spotlight (Top 5%) paper.
DemoGen	#manipulation	A demonstration generation system validated the Galaxea R1 humanoid robot. With only one real demonstration, it learns to place a banana into a basket and generalizes with high success across diverse positions and orientations. It was accepted by RSS 2025, and won the Best Long Paper Award at the SynData4CV Workshop, CVPR 2025.
DemoSpeedup	#mobile manipulation #simulation	Built on the Galaxea R1 robot, DemoSpeedup accelerates manipulation up to 3× faster. Using action entropy, it identifies "acceleration points" without manual labels, enabling efficient deployment across diverse scenarios.
General Motion Retargeting	#simulation #whole body control	Integrated with the Galaxea R1 Pro robot, GMR enables real-time high-quality whole-body retargeting for teleoperation (e.g., TWIST), supports multiple humanoid robots and human motion formats, and has been widely adopted in community research projects.
VR-Robo	#Real-to-Sim-to-Real #navigation #simulation	Validated on the Galaxea R1 robot, VR-Robo introduces a "Real-to-Sim-to-Real" paradigm for autonomous navigation. Using only RGB input—without depth or LiDAR—it enables robust real-world navigation without requiring real-robot training. It has been accepted at RA-L 2025.
Active Vision-Driven Robotic	#manipulation #simulation	Built on the Galaxea A1 robotic arm and desktop teleoperation platform, AVR achieves stable tracking and precise control with repeatability under 1 cm accuracy.
Pi0.5	#VLA	Validated on the Galaxea R1 Lite platform, π-0.5 demonstrates robust embodied intelligence by leveraging its chassis and torso to handle complex household tasks with human-like flexibility.
NavDP	#navigation #Real-to-Sim-to-Real	Powered by the Galaxea R1 robot, NavDP enables autonomous navigation in unseen real-world environments. Trained entirely in simulation, it allows robots to avoid random obstacles and safely find their way.
CogVLA	#VLA #manipulation	A cognition-aligned Vision-Language-Action framework that uses instruction-driven routing and sparsification to improve efficiency and coherence. Demonstrated on the Galaxea R1 Lite for real-world tasks and accepted at NeurIPS 2025.
Safebimanual	#manipulation	A safety-aware framework for diffusion-based bimanual manipulation. By adding test-time trajectory optimization with dynamic constraints, it prevents unsafe behaviors and improves coordinated dual-arm performance on the Galaxea R1 robot and has been accepted by CoRL 2025.
XRoboToolkit	#robot teleoperation #manipulation	XRoboToolkit is an OpenXR-based framework for robot teleoperation in extended reality. It provides low-latency feedback, precise motion control, and flexible tracking across devices. The system is fully compatible with the Galaxea R1 Lite and other platforms, enabling scalable and high-quality demonstration data collection for Vision-Language-Action models.
HERMES	#mobile manipulation	HERMES is a human-to-robot learning framework for mobile bimanual dexterous manipulation. Built on the Galaxea A1 arms and X1 mobile base, it translates human motion into physically plausible robot behaviors via reinforcement learning. With sim2real transfer and PnP-based localization, HERMES achieves robust, generalizable performance in real-world environments.
RoboBrain-X0	#mobile manipulation	A cross-embodiment foundation model from BAAI that unifies vision, language, and action for zero-shot generalization and few-shot adaptation. Deployed on the Galaxea R1 Lite for real-world task execution.
MoMaGen	#simulation #mobile manipulation	A large-scale data generation framework for bimanual mobile manipulation, addressing challenges in base placement and camera positioning for visuomotor learning. Using the Galaxea R1 robot, MoMaGen formulates data generation as a constrained optimization problem that ensures both reachability and visibility, enabling diverse datasets and robust imitation learning from a single demonstration.
DAWN	#mobile manipulation	DAWN is a novel framework for translating language instructions into robotic actions. It integrates high-level planning and low-level control into a unified diffusion process, operating within the pixel-motion space. The framework accepts language instructions (with optional visual context), generates a "pixel motion map," and converts it into robotic actions. This system has been deployed on the Galaxea R1 Lite platform. The pixel-based approach not only endows motion flows with interpretability but also enhances generalization across tasks.
iFlyBot-VLA	#VLA #mobile manipulation	iFlyBot-VLA, with Galaxea Open-World Dataset included in pre-training, adopts dual-level action representation and mixed training, achieving excellent performance in simulation and real-world manipulation.
UNICOD	#simulation #mobile manipulation	Trained on diverse datasets including Galaxea Open-World and its VQA data, integrating discrete and continuous representations via two-stage learning to achieve SOTA in simulation and real-world scenarios.
FASTER	#VLA #mobile manipulation	FASTer VLA model to optimize autoregressive VLA performance and inference efficiency; Galaxea R1 Lite hardware is used for real-world task testing to verify the model's actual operation performance.
Depth VLA	#VLA #mobile manipulation	DepthVLA integrates a depth expert, implemented on the Galaxea R1 Lite , significantly enhancing the robot's spatial reasoning and operational accuracy.
Robocoin	#dataset	RoboCOIN is currently the world’s largest and most comprehensive real-world bimanual robot manipulation dataset. It is collected across diverse real-world environments using 15 heterogeneous robot platforms, and contains 180,000+ real trajectories and 421 tasks. The dataset integrates commonly used platforms from research, industry, and education — including mobile manipulation systems such as the Galaxea R1 Lite, which were used to collect a wide range of everyday interaction behaviors.
VLA-Reasoner	#VLA	VLA-Reasoner, developed by NTU, enhances Vision-Language-Action (VLA) models for long-horizon tasks using the Galaxea A1 robotic arm. It combines trajectory rollout and Monte Carlo Tree Search (MCTS) to foresee future states and select optimal actions.
GSWorld	#sim2real	GSWorld, developed by UC San Diego and UCLA, is a photo-realistic simulator for robotic manipulation that supports the StarOcean Galaxea R1 robot. It uses the GSDF format to enable zero-shot sim2real learning, automated data collection, and reproducible benchmarking.
ENACT	#simulation	ENACT is a benchmark for evaluating the embodied cognition of VLMs. Framed within the POMDP framework, it includes two world modeling tasks: forward modeling (ordering observations based on actions) and inverse modeling (ordering actions based on observations). Leveraging the BEHAVIOR simulation, it generates 8,972 QA pairs. Using Task Accuracy and Pairwise Accuracy for evaluation.
VLASH	#manipulation	VLASH, developed by teams from MIT, NVIDIA, Tsinghua University, and UC Berkeley, is an asynchronous inference framework that adds zero overhead and requires no VLA architecture changes, while delivering smooth motion, fast response, and high accuracy.All demos use the π_0.5 model at 15 Hz on a laptop RTX 5090, with Galaxea R1 Lite as the test platform.
NORA-1.5	#manipulation	NORA-1.5, developed by researchers from Nanyang Technological University, Lambda Labs, and Singapore University of Technology and Design, is a vision-language-action model trained with world-model and action-based preference rewards and evaluated in simulation and on the Galaxea A1 robotic arm.
ViTaS	#manipulation	ViTaS, developed by researchers from Shanghai Qi Zhi Institute, Harbin Institute of Technology, Tsinghua University, The University of Hong Kong, and Carnegie Mellon University, is a visuo-tactile representation learning framework. It uses Soft Fusion Contrastive Learning and a CVAE to handle self-occluded manipulation scenarios, and was successfully deployed on the Galaxea R1.
EgoScale	#manipulation	EgoScale, developed by NVIDIA GEAR Lab, introduces a Vision-Language-Action (VLA) policy pretrained on over 20,000 hours of egocentric human video data. It utilizes an aligned human-robot mid-training stage to bridge embodiment gaps. With the Galaxea R1 Pro as its core validation platform, it achieves strong generalization for complex dexterous manipulation tasks.

🏆 Competitions & Demos

Competition/Demo Name	Tags	Description
WBCD2025-Champion	#manipulation	Using NeSyPack, the team from Carnegie Mellon University won the First Prize in the What Bimanuals Can Do (WBCD) competition at ICRA 2025, using the Galaxea A1X robotic arm.
BEHAVIOR Challenge	#dataset	Built on the Galaxea R1 Pro robot, the BEHAVIOR-1K dataset contains 50 full-length household tasks with 10,000 teleoperated demonstrations covering cooking, cleaning, rearrangement, and installation. The BEHAVIOR Challenge evaluates robots on reasoning, locomotion, and dexterous bimanual manipulation, co-hosted with NeurIPS 2025.
RoCo Challenge	#simulation #mobile manipulation	RoCo (Robotic Collaboration) is an industry-oriented benchmark initiated by NTU and presented at AAAI 2026, focusing on human-in-the-loop gearbox assembly across three core scenarios. It supports multiple robotic platforms—including the Galaxea R1 Pro—to advance adaptive collaboration and error-aware autonomy in realistic manufacturing settings.