Helix - Figure introduces an end-to-end universal control model.

What is Helix?

Helix is a general-purpose vision-language-action (VLA) model from Figure for controlling humanoid robots. Helix pioneered high-rate (200Hz) continuous control of the entire upper body of a robot (including wrists, torso, head, and fingers), supporting multi-robot collaboration where multiple robots share the same set of neural network weights to complete tasks. Helix demonstrates strong generalization ability by picking up previously unseen objects based on natural language commands. Training Helix is completely end-to-end, requiring no task-specific fine-tuning, runs on low-power GPUs, and has commercial deployment potential.

Helix's main functions

• Full upper body controlHigh-speed (200Hz) continuous control of the robot's entire upper body (including wrists, torso, head, and fingers) enables high-precision motion coordination.
• Multi-robot collaborationIt supports multiple robots running the same set of neural network weights simultaneously to collaboratively complete tasks, such as moving or organizing items together.
• Natural Language Understanding and ExecutionRobots can perform various tasks based on natural language commands, such as picking up unfamiliar items, operating drawers or refrigerators, etc.
• Strong generalization abilityIt can handle thousands of items of different shapes, sizes and materials.
• Commercial deployment capabilitiesIt runs entirely on low-power embedded GPUs, making it suitable for large-scale commercial applications.

Helix's technical principles

• System 2 (S2)An open-source Visual Language Model (VLM) based on 7B parameters is responsible for scene understanding and language understanding. It processes data at a frequency of 7-9Hz for "slow thinking" high-level objectives, transforming visual and linguistic information into semantic representations. Semantic information is encoded into continuous latent vectors and passed to System 1.
• System 1 (S1)A Transformer encoder-decoder architecture based on 80M parameters is used for low-level control. It processes at a frequency of 200Hz, enabling rapid execution and adjustment of movements. It combines the latent vectors transmitted from S2 with visual features to transform them into precise robot movements (such as wrist pose, finger control, head and torso orientation).
• End-to-end trainingThe algorithm maps raw pixels and natural language instructions to continuous action outputs and is trained using standard regression loss. A temporal offset is introduced during training to simulate the inference latency between S1 and S2, ensuring consistency between training and deployment.
• Decoupled architectureS1 and S2 operate at different time scales. S2 is responsible for high-level semantic planning, while S1 is responsible for real-time action execution. This ensures both the system's generalization ability and rapid response.
• Optimize inference deploymentOn the robot, S1 and S2 run on separate GPUs. S2 updates the latent vector asynchronously, while S1 performs motion control in real time.

Helix's project address

• Project official website:https://www.figure.ai/news/helix

Helix's technical principles

• Home servicesOrganizing items, storing things, operating home appliances, and other daily household chores.
• Multi-robot collaborationMultiple robots share a single neural network to work together to complete handling or assembly tasks.
• Item grabIt can capture previously unseen items based on natural language commands, and is suitable for logistics and warehousing.
• Industrial AutomationUsed in complex human-machine collaborative tasks, such as parts assembly and quality inspection.
• service industryProvide guidance, delivery and cleaning services in hotels, restaurants and other similar establishments.