Skip to main content

Figure 02 — The AI-First Humanoid

Duration: 45 min · Level: Foundational · Module: 1. The Humanoid Landscape · Focus: figure, VLA, hardware, deployment

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • Figure 02 stands 1.68m, weighs 60kg, and carries a 20kg payload
  • Hands feature 16 DOF each with tactile sensing on…
  • Uses a custom NVIDIA Orin-based compute stack with Heliogen…
  • OpenAI collaboration produces Vision-Language-Action (VLA) policies
  • BMW Spartanburg factory deployment (2024)

You will then consolidate these ideas in the hands-on lab below.

Why this matters​

Figure AI raised $675M at a $2.6B valuation in February 2024, backed by Microsoft, OpenAI, NVIDIA, and Amazon.

Overview​

Figure AI raised $675M at a $2.6B valuation in February 2024, backed by Microsoft, OpenAI, NVIDIA, and Amazon. Figure 02, announced August 2024, is a ground-up redesign that treats the robot as a compute platform first, mechanical system second. Its OpenAI partnership produces end-to-end neural network policies that go directly from sensor input to motor commands.

Key concepts​

Key idea

Figure 02 stands 1.68m, weighs 60kg, and carries a 20kg payload — optimized for warehouse and manufacturing tasks

  • Hands feature 16 DOF each with tactile sensing on fingertips; can pick up a USB-C cable and plug it in without external guidance
  • Uses a custom NVIDIA Orin-based compute stack with Heliogen multimodal neural network for end-to-end task execution
  • OpenAI collaboration produces Vision-Language-Action (VLA) policies: natural language instruction → sensory observation → motor output in a single forward pass
  • BMW Spartanburg factory deployment (2024): Figure 02 performing parts transfer tasks at human-comparable cycle times after ~24 hours of in-context learning
  • Figure's key architectural bet: scale data, not engineering — collect teleoperation demonstrations at scale and train general-purpose policies
Hands-on lab

Analyze Figure 02's end-to-end architecture: trace the data flow from RGB cameras → tokenized vision features → language encoder → action decoder → motor commands. Identify which components are off-the-shelf and which are custom.

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about Figure 02 stands 1.68m, weighs 60kg, and carries a 20kg payload?

optimized for warehouse and manufacturing tasks

Q2. What do you know about Hands feature 16 DOF each with tactile sensing on…?

Hands feature 16 DOF each with tactile sensing on fingertips; can pick up a USB-C cable and plug it in without external guidance

Q3. What do you know about Uses a custom NVIDIA Orin-based compute stack with Heliogen…?

Uses a custom NVIDIA Orin-based compute stack with Heliogen multimodal neural network for end-to-end task execution

Q4. What do you know about OpenAI collaboration produces Vision-Language-Action (VLA) policies?

natural language instruction → sensory observation → motor output in a single forward pass

Q5. What do you know about BMW Spartanburg factory deployment (2024)?

Figure 02 performing parts transfer tasks at human-comparable cycle times after ~24 hours of in-context learning

References​

  • Helix: A Vision-Language-Action Model for Generalist Humanoid Control — Figure AI Research (2024). Figure AI Technical Report

← Previous: 1.1 State of the Humanoid Industry · Next: 1.3 Tesla Optimus — Scale as Moat →

Part of Module 1: The Humanoid Landscape.