TechPrufer
Get started
Figure AI
Figure AI logoFigure AI · Feature trail

Helix 02 Dynamic Whole-Body Control

13 posts · Feb 2025 – Jun 2026

What this trail covers

Helix, a generalist Vision-Language-Action (VLA) model, has been enhanced for humanoid logistics and general-purpose tasks. It now handles a wider variety of package types (deformable poly bags, flat envelopes) with improved dexterity and speed (4.05s/package). Key architectural improvements include temporal vision memory for stateful perception and robustness to interruptions, integration of force feedback for a sense of touch enabling precise grips, and state history for faster, more reactive control. This latest iteration, 'Helix: A Vision-Language-Action Model for Generalist Humanoid Control', introduces full-upper-body control (35-DoF at 200Hz), multi-robot collaboration, and zero-shot generalization to novel objects via a 'System 1, System 2' VLA architecture. System 2 (VLM) handles scene understanding and language comprehension at 7-9 Hz, while System 1 (visuomotor policy) translates these into precise continuous actions at 200 Hz. The model is trained end-to-end on a multi-robot dataset, with inference optimized for onboard embedded GPUs. This enables immediate commercial deployment and significantly alters robotics' scaling trajectory by allowing new skills to be specified via natural language.

All Figure AI trails
  1. F.03 Arrives at BMWDemonstrates the first deployment of Figure 03 robot performing a logistics sequencing use case at BMW Group Plant Spartanburg, powered by Helix 02. Highlights the robot's ability to manipulate parts while stepping and repositioning its body to pull a heavy cart, showcasing dynamic whole-body control and precision/power switching within a single workflow. Details how Helix 02 enables loco-manipulation by perceiving the scene and making real-time corrections for parts that do not arrive in perfectly uniform orientations, addressing the complexities of sequencing in automotive manufacturing.
  2. Helix 02 Bedroom TidyThis post demonstrates the application of the Helix-02 VLA model to a complex collaborative task: resetting a bedroom. It showcases the model's ability to perform multi-humanoid locomanipulation directly from pixels to actions using a single learned neural network. Key advancements include opening doors with whole-body coordination, pushing furniture using stance and balance, draping clothing onto narrow fixtures, placing objects with in-hand reorientation, closing a book with dexterous bimanual control, operating a trash can foot pedal with single-leg balance, coordinating two humanoids around a shared deformable object (a comforter), and manipulating bedding with bimanual whole-body motions. The post highlights the challenges of multi-robot interaction, deformable object manipulation, and real-time decision-making in a shared environment, emphasizing that the core algorithm remained unchanged, with new data driving the expanded capabilities.
  3. Ramping Figure 03 ProductionThis post details the significant scaling of Figure 03 robot production, achieving a 24x throughput increase to one robot per hour. It highlights improvements in manufacturing quality, yield rates, and end-of-line testing. Operationally, it describes the development of robust diagnostics, fallback ladders, and a custom Fleet Management System with OTA updates. A key technical breakthrough is the integration of perception-conditioned whole-body control into Helix's System 0 (S0), enabling the AI to process visual input for environmental awareness and perform complex maneuvers like stair traversal via zero-shot sim-to-real transfer.
  4. Helix 02 Living Room TidyThis post details the expansion of Helix 02's capabilities to include whole-body, end-to-end living room cleanup. Key new behaviors demonstrated include coordinated tool use (spray bottle and towel), dynamic handling of flexible objects (towels), complex bimanual manipulation (holding a bin while scooping objects), whole-body strategies for efficiency (stowing a container), dynamic object throws (pillows), in-hand reorientation for precise tasks (remote control), temporary stowage of tools during motion, and navigation through tight spaces with precise foot placement. The post emphasizes that these new skills were learned by adding new data to the existing general-purpose architecture, without new algorithms or special-case engineering.
  5. Introducing Helix 02: Full-Body AutonomyIntroduces Helix 02, a unified whole-body loco-manipulation VLA that extends control to the entire robot. Key contributions include: 1. System 0, a learned whole-body controller trained on 1000+ hours of human motion data and sim-to-real RL, replacing 109,504 lines of hand-engineered C++ with a single neural prior for stable motion. 2. System 1's expansion to connect all sensors (head cameras, palm cameras, fingertip tactile sensors, proprioception) to all actuators (full-body joint control), enabling dexterous manipulation with new hardware modalities. 3. Demonstration of long-horizon autonomous tasks (e.g., 4-minute dishwasher loading/unloading) and new classes of dexterity (e.g., pill extraction, syringe dispensing) enabled by the integrated system.
  7. F.02 Contributed to the Production of 30,000 Cars at BMWDetails the deployment of Figure 02 at BMW for sheet-metal loading, including KPIs (cycle time, placement accuracy, interventions) and technical challenges. Highlights hardware reliability learnings, specifically the re-architecture of the forearm subsystem to improve reliability and simplify thermal management for Figure 03.
  8. Introducing Figure 03Introduced Figure 03, a third-generation humanoid robot with a ground-up hardware and software redesign. Key advancements include a redesigned sensory suite and hand system for the Helix AI, featuring a next-generation vision system with doubled frame rate, quartered latency, and a 60% wider field of view per camera, plus an embedded palm camera for redundant visual feedback. New compliant and tactile hand designs with a first-generation tactile sensor capable of detecting 3-gram forces are integrated. The robot supports 10 Gbps mmWave data offload. For home use, it incorporates safety features, an upgraded audio system, and wireless inductive charging. The design is optimized for mass manufacturing using tooled processes and a new supply chain, with the BotQ facility capable of producing 12,000 robots annually. Commercial applications benefit from faster actuators, improved torque density, and enhanced perception for precise manipulation.
  9. Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot TransferIntroduced Project Go-Big for internet-scale humanoid pretraining using a partnership with Brookfield to collect diverse real-world human video data. Achieved zero-shot human video-to-robot transfer for navigation, enabling Helix to navigate complex environments from natural language commands using only human video data. Integrated navigation capabilities into a single Helix network that also handles manipulation, eliminating the need for separate systems.
  10. Helix Loads the DishwasherThis post demonstrates that the existing Helix VLA model, previously used for folding laundry and rearranging packages, can learn to load a dishwasher with only new data. This includes singulating stacked plates, reorienting and placing glasses with precision, adapting to messy starting configurations, and recovering from errors like misgrasps or collisions. The key technical takeaway is the model's ability to acquire new, complex manipulation capabilities through data alone, highlighting the scalability of its general-purpose architecture.
  11. Helix Learns to Fold LaundryThe Helix VLA model architecture, previously applied to logistics tasks, has been successfully adapted to the complex dexterous manipulation challenge of folding laundry. This was achieved solely through data augmentation, without modifications to the model architecture or training hyperparameters. The post highlights Helix's ability to pick from mixed piles, adjust folding strategies based on initial configurations, recover from errors, and perform fine manipulation (edge tracing, pinching, unraveling) end-to-end from vision and language input to motor control. Additionally, Helix learned natural multimodal interactions such as maintaining eye contact, directing gaze, and using learned hand gestures.
  12. Scaling Helix: a New State of the Art in Humanoid LogisticsThis post details significant advancements in the Helix logistics system, focusing on architectural improvements and data scaling. New features include temporal vision memory for stateful perception and robustness, integration of force feedback for a sense of touch, and state history for faster, more reactive control. These enhancements, combined with scaling demonstration data from 10 to 60 hours, resulted in a reduction of average package handling time from ~6.84s to 4.31s (a 58% increase in throughput) and an increase in barcode scanning success rate from 88.2% to 94.4%. The post quantifies the impact of each architectural module (vision memory, state history, force feedback) through an ablation study.
  13. Helix Accelerating Real-World LogisticsThis post details improvements to Helix's System 1 (visuo-motor control policy) for real-world logistics package manipulation. Key contributions include: 1) Integration of stereo vision and a multiscale feature extraction network for richer 3D understanding and improved robustness to package variations, leading to a 60% increase in throughput over non-stereo baselines. 2) A learned visual proprioception model for online self-calibration, enabling strong cross-robot policy transfer despite hardware variations. 3) Emphasis on data quality over quantity, showing curated demonstrations yield significantly better throughput. 4) A test-time speed-up technique ('Sport Mode') that interpolates action chunk outputs, achieving up to 50% speedup with minimal performance degradation and increasing effective throughput beyond expert trajectories.
  14. Helix: A Vision-Language-Action Model for Generalist Humanoid ControlThis post introduces Helix, a Vision-Language-Action (VLA) model that achieves full-upper-body control (35-DoF at 200Hz) for humanoid robots. It employs a 'System 1, System 2' architecture where System 2 (a VLM) handles high-level scene understanding and language comprehension at 7-9 Hz, and System 1 (a visuomotor transformer) translates these into precise continuous actions at 200 Hz. This decoupled approach allows for fast, reactive control while leveraging general semantic knowledge. Key contributions include zero-shot generalization to novel objects and multi-robot collaboration, enabled by end-to-end training on a multi-robot dataset and optimized inference on embedded GPUs. The architecture avoids complex action tokenization and allows for immediate commercial deployment.