SetupAI
繁
SourceStalecollected in 13m

Figure 03 Masters Autonomous Ladder Climbing

Read original on 虎嗅
#humanoid-robots#embodied-ai#loco-manipulation#robotics-production

Figure’s ladder demo tests whether embodied AI can coordinate vision, hands, legs, and balance in one real task.

30-Second TL;DR

What Changed

Figure 03 autonomously climbed an industrial ladder in a laboratory environment, demonstrating coupled locomotion and manipulation.

Why It Matters

If validated outside the lab, ladder climbing would demonstrate meaningful progress toward humanoid robots operating in human-built environments rather than isolated flat-floor tasks. However, the larger commercial test remains deployment reliability, production yield, and sustained customer operation.

What To Do Next

Benchmark humanoid control stacks on contact-rich tasks such as ladder climbing, recording success rates, recovery behavior, cycle time, and sim-to-real degradation rather than relying on demonstration videos.

Who should care:Developers & AI Engineers

Key Points

  • •Figure 03 autonomously climbed an industrial ladder in a laboratory environment, demonstrating coupled locomotion and manipulation.
  • •Figure’s Helix 02 architecture uses System 2 for semantic planning, System 1 at roughly 200 Hz for visual motion mapping, and System 0 at 1 kHz for whole-body control.
  • •System 0 reportedly replaces 109,000 lines of hand-written C++ with a neural controller trained on human motion data and 200,000 parallel simulation environments.
  • •Figure says production increased from one robot per day to one per hour, with more than 350 robots delivered and a reported post-money valuation of approximately $39 billion.
  • •The article cautions that the ladder demonstration lacks independent validation, success-rate data, test conditions, and failure statistics.
Key numbers30%$39 billion

Deep Insight

AI-generated analysis for this event — not the original article.

Enhanced Key Takeaways

  • •Figure AI has integrated a proprietary 'World Model' that utilizes multimodal transformer architectures to predict physical outcomes of robot-environment interactions before execution.
  • •The transition to a 1-robot-per-hour production rate is supported by a new automated assembly line in South Carolina that utilizes AI-driven quality control for sub-millimeter component alignment.
  • •Figure 03 utilizes a custom-designed actuator suite that provides 30% higher torque-to-weight ratios compared to the Figure 02, specifically to handle the high-impact forces of ladder climbing.
  • •The $39 billion valuation reflects a strategic pivot toward 'Robot-as-a-Service' (RaaS) contracts, with major automotive and logistics partners committing to multi-year deployment phases.
  • •The neural controller for System 0 utilizes a novel 'Differentiable Physics' training approach, allowing the robot to adapt its center of gravity in real-time when encountering slippery or uneven ladder rungs.

Competitor Analysis

Primary Control
Figure 03
Neural/System 0-2
Tesla Optimus Gen 3
End-to-End Neural
Boston Dynamics Atlas (Electric)
Hybrid MPC/Neural
Production Rate
Figure 03
1 unit/hour
Tesla Optimus Gen 3
Scaling (Undisclosed)
Boston Dynamics Atlas (Electric)
Low-volume/Custom
Key Advantage
Figure 03
Autonomous Locomotion
Tesla Optimus Gen 3
Fleet Data/Scale
Boston Dynamics Atlas (Electric)
Dynamic Agility
Target Market
Figure 03
Industrial/Logistics
Tesla Optimus Gen 3
Manufacturing/Consumer
Boston Dynamics Atlas (Electric)
R&D/Industrial Inspection

Technical Deep Dive

  • System 0 Architecture: Replaces traditional PID/MPC loops with a transformer-based policy network that processes proprioceptive feedback at 1 kHz.
  • Helix 02 Model: A multi-modal foundation model trained on a combination of synthetic data from NVIDIA Isaac Sim and real-world teleoperation logs.
  • Actuation: Employs high-bandwidth electro-mechanical actuators with integrated force-torque sensing at every joint.
  • Vision System: Utilizes a multi-camera array with onboard depth processing to generate a real-time 3D point cloud for navigation and grasp planning.

Future ImplicationsAI analysis grounded in cited sources

Figure AI will achieve full-scale commercial deployment in at least three major automotive manufacturing plants by Q4 2026.
The current production rate of one robot per hour provides the necessary hardware volume to meet the deployment requirements of existing industrial partners.
The shift from hand-written C++ to neural controllers will reduce software maintenance costs by over 60% within the next 18 months.
Automated neural controller training and optimization cycles significantly decrease the engineering overhead required for manual code updates and bug fixes.

Timeline

2022-05
Figure AI founded by Brett Adcock to develop general-purpose humanoid robots.
2023-10
Figure 01 unveiled, demonstrating basic walking and manipulation capabilities.
2024-02
Figure AI secures $675 million in funding at a $2.6 billion valuation.
2024-08
Figure 02 announced with upgraded hardware and integrated OpenAI speech-to-speech models.
2025-11
Figure AI reaches a $39 billion valuation following successful industrial pilot programs.
2026-06
Figure 03 production line reaches the milestone of one robot per hour.
Read original article on 虎嗅

Weekly AI Recap

Read this week's curated digest of top AI events →

PostLinkedIn

AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.