Figure Helix
Figure AI's VLA Model for Humanoid Robots
Author’s Note
- Helix 02 is the first humanoid VLA to control the full body with a single neural network.
- Performing 61 sequential actions over 4 minutes without resets proves real-world capability, not cherry-picked demos.
- Impressive that they’ve pushed control frequency to 1kHz with the three-layer System 0/1/2 architecture.
Key Significance
- First Full-Body Autonomous Humanoid: Single neural system that controls the full body directly from pixels
- Longest Autonomous Task: 4-minute continuous operation with 61 sequential loco-manipulation actions, no resets or human intervention
- Three-Layer Architecture: System 0 (1kHz) + System 1 (200Hz) + System 2 (semantic reasoning)
- New Sensors: Palm cameras and tactile sensors detecting forces as small as 3g
- Replaces 100K Lines: 10-million parameter neural network replaces 109,504 lines of hand-engineered C++ code
Figure Helix 02 Official Demo Video (2026.01)
Overview
Helix 02 (2026.01)
Key Advancements
Helix 02 is the first humanoid with “a single neural system that controls the full body directly from pixels.”
| Feature | Helix (2025.02) | Helix 02 (2026.01) |
|---|
| Control Range | Upper body focused | Full body (locomotion + manipulation integrated) |
| Architecture | System 1/2 | System 0/1/2 |
| Max Frequency | 200Hz | 1kHz |
| Tactile Sensors | None | Detects 3g forces |
| Palm Cameras | None | Yes |
4-Minute Continuous Autonomous Task
Dishwasher loading/unloading across an entire kitchen:
| Item | Value |
|---|
| Continuous Operation Time | 4 minutes |
| Sequential Loco-manipulation Actions | 61 |
| Resets/Human Intervention | None |
Dexterity Demonstrations
- Unscrewing bottle caps with torque control
- Extracting individual pills from organizers despite occlusion
- Dispensing precisely 5ml from syringes
- Singulating small metal pieces from cluttered bins
Architecture: System 0/1/2
Helix 02 uses a three-layer hierarchical architecture.
Helix 02 System 0/1/2 Architecture Explanation
System 0 (S0) - Physical Execution Layer
| Item | Details |
|---|
| Frequency | 1 kHz |
| Parameters | 10 million |
| Training Data | 1,000+ hours of human motion data |
| Simulation | 200,000+ parallel environments with domain randomization |
| Replaced Code | 109,504 lines of hand-engineered C++ |
Handles real-time balance and contact management.
System 1 (S1) - Visuomotor Control
| Item | Details |
|---|
| Frequency | 200 Hz |
| Architecture | Transformer conditioned on System 2 latents |
| Structure | ”Pixels-to-whole-body” (all sensors → all actuators) |
Sensor Inputs:
- Head cameras
- Palm cameras
- Fingertip tactile sensors
- Proprioception
Actuator Outputs:
- Complete joint control: legs, torso, head, arms, wrists, individual fingers
System 2 (S2) - Semantic Reasoning
- Processes visual scenes and language commands
- Sequences high-level behaviors (“Walk to dishwasher and open it”)
- Generates latent goals without specifying low-level coordination
New Hardware Capabilities
Palm Cameras
- Cameras mounted on palms
- Enable in-hand visual feedback when objects are occluded from head perspective
Tactile Sensors
- Sensitivity: Detects forces as small as 3g (“sensitive enough to feel a paperclip”)
- Force-modulated grasping across five-fingered hands
Helix (2025.02) - Initial Version
Helix Architecture: System 1 (200Hz low-level control) + System 2 (7-9Hz high-level planning)
Key Features
- First full-body high-speed control VLA for humanoids
- Controls upper body (wrists, torso, head, individual fingers) at 200Hz
- Dual robot simultaneous control
Table-to-Dishwasher Task
| Item | Value |
|---|
| Distance Traveled | 130+ feet |
| Unique Interactions | 33 |
| Number of Objects | 21 (including delicate dishes) |
| Item | Spec |
|---|
| DoF | 35 DoF |
| Hands | Human-like wrists, hands, fingers |
| Computing | NVIDIA RTX GPU (3x previous) |
| Cameras | Head 6x RGB + Palm cameras (Helix 02) |
| Tactile Sensors | Fingertips (Helix 02) |
| Hand Payload | Up to 25kg |
Figure 03 is the third-generation humanoid robot designed for home environments.
| Feature | Figure 02 | Figure 03 |
|---|
| Target Environment | Industrial (BMW factory) | Home |
| Weight | Baseline | 9% lighter |
| Exterior | Hard machined parts | Soft textiles + multi-density foam |
| Battery | - | 5-hour runtime, 2kW wireless charging |
| Data Transfer | - | 10 Gbps mmWave |
Vision System
| Item | Spec |
|---|
| Total Cameras | 8 (Head 6 + Palm 2) |
| Frame Rate | 2x improved |
| Latency | Reduced to 1/4 |
| Field of View | 60% wider |
Tactile Sensors
- Fingertip sensors: Detect forces as small as 3g (paperclip weight)
- Custom-developed first-generation tactile sensors
Battery & Charging
| Item | Spec |
|---|
| Model | F.03 Battery |
| Runtime | ~5 hours |
| Charging | 2kW wireless inductive charging (foot coils) |
| Certification | UN38.3, targeting UL standard |
| Data Offload | 10 Gbps mmWave (for fleet learning) |
Safety Design
- Covered in soft textiles (washable, tool-free removal)
- Multi-density foam eliminates pinch points
- Optional: Cut-resistant garments
- Customizable side screens (fleet branding)
Manufacturing: BotQ
| Item | Details |
|---|
| Current Capacity | 12,000 units/year |
| 4-Year Goal | 100,000 total units |
| Manufacturing | Shift from CNC to die-casting, injection molding, stamping |
| Expected Price | ~$20,000 |
Figure AI History
| Date | Event |
|---|
| 2022 | Figure 01 - First prototype, bipedal, targeting logistics/warehousing |
| 2024.08 | Figure 02 - Industrial deployment, BMW factory testing |
| 2025.02 | Helix - First full-body humanoid VLA |
| 2025.10 | Figure 03 - Home humanoid, soft design |
| 2026.01 | Helix 02 - Full-body autonomy, System 0/1/2 architecture |
Funding & Partnerships
| Timing | Details |
|---|
| 2024.01 | BMW partnership (automotive manufacturing deployment) |
| 2024.02 | $675M funding (valuation $2.6B) |
| Investors | Jeff Bezos, Microsoft, NVIDIA, Intel, Amazon, OpenAI |
References
See Also