Reborn Model Zoo

What Models Does Reborn Provide?

Reborn offers a growing library of pretrained and adaptable embodied intelligence models. Here we list the main streams that can be adopted in most existing humanoid robots:

• Versatile Vision-Language-Action (VLA) Models Translate high-level language instructions and visual inputs into action sequences (e.g., “Pick up the red bottle and place it on the table”).

• Full-body Humanoid Control Models Trained using mocap and VR data to power humanoid locomotion, balance, posture control, and goal-directed motion.

• Dexterous Manipulation Models Learned from VR and first-person videos to handle fine-grained in-hand tasks, such as grasping, rotating, and tool use.

The Reborn model zoo contains more than these models, which supports one-click deployment on real robots, enabling practical tasks such as home assistance, object retrieval, or interactive behaviors in educational and service settings.

The Reborn Model Zoo

👁 OpenVLA

Vision-Language-Action (VLA) Model

OpenVLA enables robots to interpret natural language instructions and execute multi-step plans in visually rich environments. Trained on Reborn’s ego-centric video and simulation data, this model bridges perception and intent, allowing robots to respond to prompts like “place the red cup on the shelf” with contextual understanding and spatial awareness.

🕴 Humanoid Whole-Body Control

Locomotion + Balance + Posture

Designed for full-scale humanoid platforms, this model translates task-level commands into low-level whole-body movement. It is trained using motion capture sequences and VR locomotion data, enabling smooth walking, turning, bending, and balancing behaviors for platforms like Unitree G1 and Figure 02.

✋ Open-Vocabulary Pick and Place

Grasp Anything, Anywhere

This model allows robots to pick and place objects based on open-vocabulary inputs such as “pick up the toothbrush” or “move the snack box to the left table.” It integrates VLA reasoning with manipulation execution, supporting flexible object recognition, grasp planning, and placement in unstructured environments.

🤚 Open-Vocabulary Articulated Manipulation

Tools, Drawers, Doors, and More

Beyond simple pick-and-place, this model specializes in interacting with articulated objects—e.g., opening cabinets, operating switches, or using utensils. It combines semantic intent parsing with force-aware control, enabling robots to perform complex object interactions safely and accurately.

🧤 Dexterous Grasping

In-Hand Manipulation at Human Level

Trained with high-resolution hand motion from VR devices, this model controls robotic hands with fine-grained precision. It enables capabilities like rotating an object inside the palm, regripping tools, or handling soft or irregular objects, unlocking human-like dexterity for next-gen robot hands.