Figure’s Helix: Advancing Humanoid Robotics in Logistics

Figure is bringing humanoid robots into the workforce, starting with logistics. Handling packages may seem simple, but it requires speed, precision, and adaptability—challenges that push robotics to new limits. With advancements in vision, movement, and learning, Figure’s robots are mastering real-world tasks with human-like dexterity. Their latest innovations even allow robots to work faster than their trainers, proving that AI-powered automation is ready for high-speed, high-precision industries. But this is just the beginning. As humanoid robots take on more complex roles, Figure is shaping the future of work. How far can robotics go? The answer is unfolding now.

Helix’s Vision-Language-Action Model

Helix is Figure’s advanced Vision-Language-Action (VLA) model, designed to give humanoid robots a deeper understanding of the world around them. It combines perception, language comprehension, and learned control, allowing robots to interpret their environment, understand commands, and execute tasks with human-like dexterity.

A core component of Helix is System 1 (S1)—its low-level visuo-motor control policy. This system refines how robots see, move, and manipulate objects, making tasks like package handling in logistics faster and more precise. With Helix, robots can adapt, self-correct, and perform complex movements, bringing AI-powered automation closer to real-world applications.

Key Advancements in System 1:

Helix’s System 1 (S1) has undergone significant improvements, enhancing its ability to perceive, adapt, and execute complex tasks with greater precision and efficiency. These advancements enable Figure’s humanoid robots to handle real-world logistics challenges more effectively while ensuring seamless scalability across multiple robots.

Implicit Stereo Vision:

Helix now has a richer 3D understanding of its environment, allowing it to accurately perceive depth and spatial relationships. Unlike traditional monocular vision, which struggles with distance estimation, implicit stereo visionenables robots to assess object size, shape, and position with greater precision. This enhancement improves grasping accuracy, movement planning, and adaptability to unpredictable environments, making tasks like package manipulation more reliable and efficient.

Multi-Scale Visual Representation:

Helix processes visual information using a multi-scale approach, capturing both fine-grained details and broader scene context. This means it can recognize tiny features, like a shipping label’s orientation, while still understanding larger environmental cues, such as conveyor belt movement. This dual-layered perception enables better object handling, smoother decision-making, and faster adaptation to new settings, ensuring high performance across different logistics scenarios.

Learned Visual Proprioception:

Each Helix-powered robot is now capable of self-calibration, making cross-robot deployment seamless. Instead of relying on manual tuning—which can be time-consuming and inconsistent—Helix learns to interpret its own physical state through vision-based proprioception. This advancement reduces performance inconsistencies caused by hardware differences between individual robots, improving scalability and fleet-wide efficiency without requiring per-robot adjustments.

Sport Mode:

To enhance speed and efficiency, Helix now features Sport Mode, a test-time speed-up technique that enables robots to execute tasks faster than their human demonstrators—without sacrificing accuracy or dexterity. By intelligently resampling movement trajectories, Sport Mode allows for up to a 50% increase in execution speed, making package handling more efficient and high-throughput. This ensures robots can keep up with demanding industrial workflows while maintaining precise and controlled movements.

These advancements collectively push the limits of AI-driven robotics, allowing Figure’s humanoid robots to perform real-world logistics tasks at unprecedented speed, precision, and reliability.

Real-World Application: Logistics Package Handling:

Helix has been rigorously tested in real-world logistics environments, where efficiency, precision, and adaptability are critical. One of its primary tasks involves transferring packages from one conveyor belt to another while ensuring that shipping labels are correctly oriented for scanning. Although this may seem straightforward, it presents several complex challenges that demand human-level dexterity and decision-making.

Challenges in Package Handling:

1. Variability in Package Properties:

   • Packages come in different sizes, shapes, weights, and materials—from rigid boxes to flexible bags and even deformable plastic mailers.
   • The system must adjust its grasping technique in real time, ensuring a secure hold regardless of package rigidity.

2. Dynamic Conveyor Belt Movement:

   • Unlike stationary objects, packages move continuously and unpredictably on conveyor belts.
   • Helix must track objects in motion, predicting the best moment to grasp them while avoiding collisions with other items.

3. Label Orientation and Precision Placement:

   • For automated scanning systems to work efficiently, shipping labels must be properly oriented when packages are placed on the destination conveyor.
   • Helix must detect and rotate packages accordingly, even if labels are partially obscured or difficult to read.

How Helix Overcomes These Challenges:

Thanks to its recent System 1 (S1) improvements, Helix now performs these tasks with unmatched efficiency and adaptability:

• Enhanced Stereo Vision for Smarter Grasping:

  • Helix’s new implicit stereo vision provides a rich 3D understanding of its environment.
  • This allows the robot to accurately judge depth and identify optimal grasp points, even when handling irregularly shaped or unfamiliar objects.

• Self-Calibration for Seamless Adjustments:

  • Helix’s learned visual proprioception enables it to self-calibrate, meaning each robot can adapt to slight differences in hardware or sensor alignment.
  • This ensures consistent performance across multiple robots without requiring individual manual tuning.

• Increased Throughput and Expanded Object Handling:

  • The combination of stereo vision and improved multi-scale perception has led to a 60% increase in package throughput compared to previous, non-stereo models.
  • Helix has also demonstrated the ability to handle previously unseen objects, including flat envelopes, which pose unique challenges due to their thin and flexible nature.

With these advancements, Helix is transforming logistics automation, making it possible for humanoid robots to match—and even surpass—human efficiency in package handling. This marks a major step toward scalable, AI-driven workforce integration in high-speed, high-precision industries.

Optimized Data Utilization:

Traditional AI models often rely on massive datasets to improve performance, but Helix takes a different, more efficient approach. Instead of requiring huge amounts of raw training data, Helix demonstrates that high-quality, well-curated datasets can yield superior results with far less data.

Why Data Quality Matters More Than Quantity:

1. More Focused Learning

   • Instead of overwhelming the system with millions of low-quality demonstrations, Helix is trained on select, high-quality human demonstrations that emphasize successful execution, adaptive corrections, and efficient motion strategies.
   • This allows the model to learn precise, dexterous behaviors faster, without being diluted by irrelevant or suboptimal examples.

2. Efficient Use of Demonstration Data

   • Just eight hours of carefully curated human demonstration data was enough to train Helix to manipulate packages with expert-level dexterity.
   • The model prioritizes learning from the best examples, ensuring it adopts optimal strategies while avoiding errors from poor demonstrations.

3. Better Generalization to New Scenarios

   • Helix’s data strategy ensures it can adapt to new environments more effectively than models trained on vast but lower-quality datasets.
   • By focusing on core manipulation principles rather than memorizing excessive variations, Helix can generalize its skills to previously unseen package types, shapes, and motion patterns.

This optimized data utilization proves that AI-powered robots don’t need enormous amounts of data to achieve expert performance—they just need the right data.

Scalability and Cross-Robot Adaptability:

One of Helix’s most groundbreaking features is its ability to seamlessly transfer learned behaviors across multiple robots. In traditional robotic systems, each unit often requires individual calibration due to slight variations in hardware, sensors, and mechanical response. Helix eliminates this challenge, making large-scale deployment significantly more practical.

How Helix Enables Cross-Robot Adaptability:

1. Online Self-Calibration

   • Helix robots feature a learned visual proprioception system, allowing them to calibrate themselves in real time.
   • Instead of relying on manual adjustments—which are time-consuming and error-prone—each robot automatically fine-tunes its perception and movement, ensuring consistent performance across different units.

2. Mitigating Hardware Variations

   • No two robots are exactly the same—small differences in sensor calibration, joint response, and camera alignment can create significant discrepancies in performance.
   • Helix compensates for these differences, ensuring that a policy trained on one robot can seamlessly transfer to others without a loss in precision.

3. Effortless Fleet-Wide Deployment

   • Traditional robotic deployments require extensive per-unit setup, making large-scale automation costly and slow.
   • With Helix’s self-calibration, Figure’s robots can be deployed at scale with minimal manual intervention, reducing operational downtime and increasing efficiency.

By enabling smooth cross-robot adaptation, Helix brings humanoid robotics one step closer to scalable, real-world workforce integration—paving the way for large-scale AI-driven automation in industries where precision and efficiency are paramount.

Conclusion:

Helix represents a major step toward integrating humanoid robots into industrial applications. With its advanced vision system, rapid learning capabilities, and adaptability, it is proving that humanoid robots can match—and even surpass—human efficiency in logistics operations. As Figure continues to push the boundaries of embodied AI, the potential for humanoid robots in the workforce is becoming a reality.