Robots: Figure 02 and Tesla Optimus Progress

Embodied AI is finally getting real. Figure’s second-generation robot and Tesla’s Optimus are making tangible progress. The era of useful humanoid robots is approaching—slowly, with caveats.

Figure 02

What It Is

Figure 02 is the second-generation humanoid from Figure AI, backed by OpenAI and NVIDIA:

OpenAI Integration

Figure 02 uses OpenAI models for reasoning:

User: "Can you make me a coffee?"
Robot: [Processes natural language]
       [Plans task sequence]
       [Executes physically]

The integration enables:

• Natural language task instructions
• Visual scene understanding
• Adaptive task execution
• Explanation of actions

Current Capabilities

Demonstrated tasks:

• Packing items in boxes
• Operating coffee machines
• Sorting objects
• Navigating warehouse environments
• Simple assembly tasks

Deployment

Figure partnered with BMW for factory trials:

• Material handling
• Quality inspection
• Repetitive assembly
• Human collaboration tasks

Tesla Optimus (Gen 2)

Specs

Progress

Tesla’s approach differs from Figure:

• Fully in-house development
• Tesla’s Dojo compute for training
• FSD vision system adapted
• Manufacturing focus from day one

Demonstrated Tasks

• Folding laundry
• Sorting objects
• Walking on various terrains
• Squats and balance exercises
• Egg handling (delicate manipulation)

Factory Deployment

Already working in Tesla factories:

• Battery cell sorting
• Limited assembly tasks
• Moving components

The Technology Stack

Sensors

┌─────────────────────────────────┐
│           Perception            │
├─────────────────────────────────┤
│ Cameras (stereo vision)         │
│ LiDAR (some models)             │
│ Force/torque sensors            │
│ IMU (balance)                   │
│ Tactile sensors (hands)         │
└─────────────────────────────────┘

Computing

┌─────────────────────────────────┐
│           On-board              │
├─────────────────────────────────┤
│ Vision processing: NVIDIA GPUs  │
│ Motion control: Real-time CPU   │
│ LLM inference: Edge chips       │
│ Cloud: Training only            │
└─────────────────────────────────┘

Training

Real robot data + Simulation + Internet data = Robot capabilities

Key insight: LLMs provide the “brain,” but motor control requires separate training on physical data.

Challenges

The Reality Gap

Simulation to reality transfer is imperfect:

Simulated task success: 95%
Real-world task success: 60-80%

Every physical detail matters: friction, compliance, object variability.

Manipulation

Human hands are incredibly sophisticated:

Human hand: 27 bones, 34 muscles, thousands of tactile sensors
Robot hand: Best cases have 20 DOF, limited sensing

Dexterous manipulation remains the hardest problem.

Cost

Volume production needed to hit targets.

Safety

Robots strong enough to be useful are dangerous:

# Oversimplified safety logic
if contact_force > safe_threshold:
    stop_immediately()
    back_away()

Real safety is much more complex—especially around humans.

Use Cases (Realistic, 2025)

Factory Work

• Repetitive manipulation
• Material transport
• Quality inspection
• Assembly assistance

Logistics

• Warehouse picking
• Package handling
• Truck loading/unloading

Healthcare (Longer Term)

• Patient assistance
• Delivery in hospitals
• Routine tasks

Home (Even Longer Term)

• Laundry
• Cleaning
• Cooking
• Elder care

What This Means for Developers

New Platforms Emerging

# Robot Operating System (ROS) 2
import rclpy
from geometry_msgs.msg import Twist

def move_robot():
    node = rclpy.create_node('robot_mover')
    publisher = node.create_publisher(Twist, '/cmd_vel', 10)
    
    msg = Twist()
    msg.linear.x = 0.5
    publisher.publish(msg)

LLM + Robotics APIs

# Conceptual high-level control
async def make_coffee(robot, llm):
    plan = await llm.generate(
        "Describe steps to make coffee in this kitchen",
        context=robot.scene_understanding()
    )
    
    for step in plan.steps:
        primitive = await llm.translate_to_robot_action(step)
        await robot.execute(primitive)

Skills to Learn

• Robot Operating System (ROS 2)
• Reinforcement learning for control
• Computer vision for robotics
• Simulation (Isaac Sim, MuJoCo)

Timeline Expectations

Take Tesla’s timelines and double them. Always.

Final Thoughts

Humanoid robots are real and improving. But we’re in the “early internet” phase—impressive demos, limited utility.

Watch Figure and Tesla. Learn robotics fundamentals. The field is accelerating.

The robots are coming. They’re just late and clumsy—for now.

The future walks on two legs. Slowly. And sometimes falls.