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

Humanoid Locomotion Control

2 posts · Feb 2025 – Sep 2026

What this trail covers

Helix 2.5, the latest iteration of Figure AI's neural network, demonstrates significant advancements in zero-shot generalization for humanoid robots. Building upon Helix 02's whole-body control capabilities, Helix 2.5 leverages pretraining on the 'Index' dataset to enable robots to perform complex tasks like tidying rooms, folding towels, and making beds in 30 previously unseen homes without any prior data collection or fine-tuning in those environments. This represents a major step towards robots understanding and adapting to new physical environments autonomously. The system shows improved efficiency in behavior specification and exhibits a human-to-humanoid robot transfer scaling law, suggesting predictable improvements with increased pretraining data.

All Figure AI trails
  1. Helix 2.5: Zero-Shot 30-Home GeneralizationHelix 2.5 introduces zero-shot whole-body generalization across 30 unseen homes for three distinct tasks (tidying, folding, bed-making). This capability is primarily driven by pretraining on the 'Index' dataset, which accounts for a 6x increase in zero-shot success rate compared to models trained from scratch (9% vs. 56%). The post quantifies this by demonstrating a human-to-humanoid robot transfer scaling law, showing predictable improvements in downstream action prediction with increased pretraining data. It also highlights that Helix 2.5 requires 2x less adaptation data than previous models for similar or broader generalized behaviors and exhibits enhanced self-correction capabilities.
  2. F.03 Arrives at BMWFigure 03, powered by Helix 02, is demonstrated performing a logistics workflow at BMW Group Plant Spartanburg, specifically a sequencing use case. This involves dynamic whole-body control where the robot manipulates parts while stepping and repositioning its body to pull a heavy cart. Helix 02 coordinates the robot's limbs and torso for loco-manipulation, continuously adapting movements to spatial variation and correcting errors. This showcases the ability to automate dynamic material manipulation and handle parts that do not arrive in mathematically perfect orientations, a task intractable for fixed automation.
  3. Helix 02 Bedroom TidyThis post demonstrates a major step in multi-humanoid collaborative locomanipulation using a single learned Vision-Language-Action (VLA) policy. Two Helix-02 equipped humanoids successfully reset a bedroom in under two minutes, performing complex tasks such as opening doors, pushing furniture, draping clothing, placing objects with reorientation, closing a book, operating a trash can pedal, and collaboratively manipulating a bed comforter. The key technical challenge addressed is the emergent complexity of two humanoids acting in a shared, dynamic environment where each robot infers the other's intent from motion alone, without explicit communication. The system handles deformable objects and requires thousands of consecutive correct decisions at policy rate, conditioned on a fast-moving scene including a second humanoid. This is presented as the first demonstration of a single learned neural network performing multi-humanoid collaborative locomanipulation directly from pixels to actions.
  4. Helix 02 Living Room TidyThis post details the expansion of Helix 02's capabilities to include 'Living Room Tidy'. The system learned to perform complex tasks such as coordinated tool use (spray bottle and towel), dynamic handling of flexible objects (towel), bimanual manipulation (scooping blocks into a bin), whole-body strategies for efficiency (tucking a container under an arm), dynamic object throws (pillow), in-hand reorientation for precise tasks (remote control), temporary stowage of tools during motion, and navigation through tight spaces with precise foot placement. The key technical delta is the demonstration that Helix 02 can learn new, complex tasks requiring integrated locomotion, dexterity, and sensing solely by adding new data, without requiring new algorithms or special-case engineering.
  5. Introducing Helix 02: Full-Body AutonomyIntroduces Helix 02, a unified whole-body loco-manipulation VLA model. Key contributions include: System 0, a 10M-parameter neural network trained on 1000+ hours of human motion data for whole-body control at 1kHz; System 1, a transformer policy that now takes all sensors (head cameras, palm cameras, tactile sensors, proprioception) as input and outputs full-body joint targets; and System 2's expanded scope for longer-horizon task sequencing. Demonstrates autonomous 4-minute dishwasher loading/unloading and dexterous manipulation tasks (unscrewing bottle caps, pill extraction, syringe dispensing, picking metal pieces) enabled by new tactile and in-hand vision sensing.
  7. Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot TransferIntroduces Project Go-Big for internet-scale humanoid pretraining using human video data from Brookfield properties. Achieves zero-shot human video-to-robot transfer for navigation, enabling robots to respond to natural language commands like 'go to the fridge' after training exclusively on egocentric human video. A single Helix neural network now outputs both high-rate dexterous manipulation and navigation commands, eliminating the need for separate systems.
  8. Natural Humanoid Walk Using Reinforcement LearningThis post details the application of end-to-end reinforcement learning (RL) trained in simulation to achieve natural humanoid walking for the Figure 02 robot. It highlights the use of high-fidelity physics simulation with domain randomization and kHz-rate torque feedback for zero-shot sim-to-real transfer. The approach incorporates reward shaping to mimic human walking gait attributes (heel-strikes, toe-offs, arm-swing) while optimizing for velocity tracking, power consumption, and robustness to perturbations. The engineering effort focuses on bridging the sim-to-real gap to enable repeatable, human-like walking across the robot fleet.
  9. Helix: A Vision-Language-Action Model for Generalist Humanoid ControlThis post introduces Helix, a Vision-Language-Action (VLA) model that enables generalist humanoid control. Key technical contributions include: 1. Full-upper-body control (35-DoF at 200Hz) with individual finger, wrist, torso, and head coordination. 2. Multi-robot collaboration capabilities. 3. Zero-shot generalization to pick up novel objects via natural language prompts. 4. A unified neural network architecture (single set of weights) for diverse behaviors without task-specific fine-tuning. 5. Onboard embedded GPU deployment for commercial readiness. The core architecture is a 'System 1, System 2' VLA: System 2 (VLM) for high-level understanding and System 1 (visuomotor policy) for low-level, high-rate control. Training uses a novel auto-labeling VLM for instruction generation and end-to-end optimization. Inference is optimized for model parallel deployment on dual embedded GPUs.