AI in Autonomy and Robotics

Learning Objectives

• Explain the AI stack that enables autonomous vehicles — perception, prediction, and planning
• Compare the leading autonomous vehicle programs and their commercial status
• Understand the state of humanoid robotics and why physical AI is harder than language AI

Autonomous Vehicles: The Longest AI Bet

Autonomous vehicles represent one of the most ambitious AI engineering challenges of the past decade. The original promise — fully self-driving cars available by 2020 — proved wildly optimistic. But the underlying AI technology has advanced dramatically, and commercial deployment is now a reality in specific, constrained geographies.

The AI challenge in autonomous driving is not any single problem but a stack of interconnected ones:

• Perception: Identifying every relevant object in the environment — other vehicles, pedestrians, cyclists, road markings, traffic signals, construction zones — in real time from sensor data (cameras, LiDAR, radar)
• Prediction: Forecasting how every detected object will behave in the next several seconds — not just where they are, but where they will be
• Planning: Computing a safe, efficient path through the predicted future environment — and replanning continuously as conditions change
• Edge cases: The "long tail" of rare but dangerous situations that trained models have seen rarely or never — a child running into the road, a mattress falling off a truck, a malfunctioning traffic signal

The long tail problem is why autonomous driving proved harder than expected: a model that handles 99.9% of situations correctly still fails thousands of times per million miles.

Waymo: The Commercial Leader

Waymo (a Google/Alphabet company, founded as the Google Self-Driving Car Project in 2009) is the most commercially advanced autonomous vehicle program in the world.

Waymo's Waymo One robotaxi service now operates in 10 US cities — Phoenix, San Francisco, Los Angeles, Austin, Dallas, Houston, San Antonio, Orlando, Atlanta, and Miami — providing paid autonomous rides to the public with no human driver in the vehicle. As of early 2026, Waymo has logged over 170 million fully autonomous miles (driving 4+ million miles per week) with a safety record of 91% fewer serious-injury crashes and 80% fewer injury crashes compared to human drivers. Swiss Re insurance data confirms 92% fewer bodily injury claims and 88% fewer property damage claims. In February 2026, Waymo raised $16 billion at a $126 billion valuation and is on track for 1 million rides per week by end of 2026, with international expansion planned for London and testing underway in Tokyo.

Waymo's technical approach:

• Multi-sensor fusion: Cameras, LiDAR, and radar provide redundant sensing with different strengths in different conditions
• High-definition maps: Pre-built detailed maps of operational areas that provide context the sensors alone cannot
• Conservative operational design domain: Waymo operates only in specific geographic areas with known conditions, allowing it to optimize for that environment
• Fleet learning: Incidents and edge cases from the entire fleet improve models for all vehicles

The limitation of Waymo's approach is scaling — the high-definition mapping requirement means expansion to new cities requires significant preparation. Plans for 20+ cities in 2026 suggest the company is accelerating this process. That said, safety incidents still occur: NHTSA opened an investigation after a Waymo struck a child in a Santa Monica school zone in January 2026 (the child sustained minor injuries; the vehicle had braked to 6 mph), and Waymo recalled 3,000+ vehicles after robotaxis illegally passed stopped school buses in Austin 19 times.

Tesla Full Self-Driving: The Scale Approach

Tesla's Full Self-Driving (FSD) takes a fundamentally different approach: pure computer vision (cameras only, no LiDAR) and end-to-end neural networks trained on video from the entire Tesla fleet.

Tesla has access to data from millions of vehicles driving billions of miles per year — an enormous training dataset advantage. Tesla's AI team (now also training for Optimus robots) has made significant progress: FSD has improved from a liability-generating feature to a genuinely capable driver-assistance system that handles highway driving, urban navigation, and parking in most conditions.

The critical distinction: Tesla FSD is a driver assistance system — it requires an attentive human driver ready to intervene. Waymo's robotaxi service is fully autonomous (SAE Level 4). These are meaningfully different products.

Tesla has produced the first Cybercab at Giga Texas — a fully autonomous robotaxi without a steering wheel or pedals — with volume production confirmed to begin in April 2026. Whether FSD's vision-only approach will achieve Level 4 safety without the redundant sensing of LiDAR-equipped competitors remains the central debate in autonomous vehicle development.

Aurora Innovation: Autonomous Trucking

Aurora Innovation is applying autonomous vehicle technology to long-haul trucking — specifically, interstate highway driving where the environment is more predictable and the business case (eliminating driver wages on 500-mile routes) is clearer.

Aurora's technology focuses on highway driving in defined corridors, avoiding the complex urban environments that make robotaxis difficult. Since launching driverless operations in April 2025, Aurora has logged 250,000+ incident-free driverless miles across Texas, New Mexico, and Arizona (including a Fort Worth to El Paso corridor). Aurora targets 200+ trucks operating by end of 2026, with next-generation hardware that doubles LiDAR range to 1,000 meters and halves hardware costs. Plans to haul freight without any observer in the cab are targeted for Q2 2026.

Zoox (Amazon): Autonomous Ride-Hailing

Zoox (acquired by Amazon in 2020) has built a purpose-built autonomous vehicle — a bidirectional vehicle without a traditional front or back, designed specifically for autonomous operation rather than retrofitting autonomy onto a human-driven platform.

Zoox launched driverless robotaxi service in Las Vegas (September 2025) and San Francisco (November 2025), with approximately 50 robotaxis in service serving 300,000+ riders and completing 450,000 paid rides per week. In March 2026, Zoox announced a partnership with Uber — Zoox vehicles will be hailable via the Uber app in Las Vegas (summer 2026) and Los Angeles (2027). Zoox is also testing in Phoenix and Dallas, with Amazon's resources enabling rapid scaling.

Humanoid Robotics: Physical AI at the Frontier

The convergence of advanced AI with robotics is creating a new category: general-purpose humanoid robots capable of performing physical tasks in human-designed environments.

The engineering challenge: robotics has historically been a specialized domain where each robot was precisely engineered for one specific task in a controlled environment (automotive assembly robots, warehouse picking robots). The humanoid vision is a robot that can operate in unstructured environments and perform diverse physical tasks — more like a human assistant than a specialized machine.

Figure AI and the Race to General Purpose Robotics

Figure AI (founded 2022, raised over $1.9 billion at a $39 billion valuation from Microsoft, OpenAI, Nvidia, and others) has already completed its first commercial deployment. Figure 02 operated at BMW's Spartanburg plant for 11 months, assisting in the production of over 30,000 BMW X3s across 1,250+ operational hours. Figure 02 is now being retired fleet-wide in favor of the next generation.

Figure 03 (announced October 2025) is designed for home use at a target price of $20,000 — standing 5'6", weighing 61 kg, with a 20 kg payload capacity, 5-hour battery, wireless charging, and fingertip sensors that detect forces as small as 3 grams. Limited home deployments are targeted for late 2026.

OpenAI's involvement is notable: OpenAI provided the language model that allows Figure robots to understand and respond to natural language instructions, enabling a human to direct the robot verbally rather than through programming.

Boston Dynamics: The Established Leader

Boston Dynamics (founded 1992, now owned by Hyundai) is the most established robotics company, famous for its viral videos of humanoid (Atlas) and quadruped (Spot) robots performing impressive physical feats.

Spot is commercially deployed as an industrial inspection robot — autonomously patrolling facilities, collecting sensor data, and performing visual inspection in environments hazardous for humans. Spot is in commercial use at construction sites, oil and gas facilities, and manufacturing plants.

Atlas (the humanoid platform) transitioned from hydraulic to electric actuation and was unveiled in production form at CES 2026 — with 56 degrees of freedom, fully rotational joints, a 2.3-meter reach, and the ability to lift 50 kg. Atlas is now commercially deployed, with production units shipping to Hyundai's Robotics Metaplant Application Center and Google DeepMind (for integration with Gemini Robotics foundation models). All 2026 deployments are fully committed, with a 30,000-unit/year factory planned for 2028.

Tesla Optimus: Manufacturing Scale as the Advantage

Tesla's Optimus program leverages Tesla's unique advantages: the AI infrastructure built for Autopilot/FSD, the manufacturing capability to produce hardware at scale, and the operational context of Tesla's own factories.

Tesla is training Optimus robots using the same video-based learning infrastructure developed for autonomous vehicles. Gen 2 robots have been deployed internally at Fremont and Austin factories since mid-2024, performing battery cell sorting, parts handling, and quality inspection. Gen 3 mass production began January 2026 at Fremont, featuring hands with 25 actuators each (50 total, a 4.5x increase from Gen 2). Tesla CEO Elon Musk clarified on the Q4 2025 earnings call that Gen 3 robots are currently for "learning and data collection" — not yet doing fully productive work.

Tesla maintains a consumer price target under $20,000 per unit at scale, with limited external sales targeted for late 2026 and broader availability in 2027. Fremont's Model S/X production lines are being converted to an Optimus manufacturing hub targeting 1 million units per year.

Unitree: Chinese Humanoid Robotics

Unitree Robotics (China) produces some of the most cost-competitive humanoid and quadruped robots available. The Unitree G1 humanoid robot starts at $13,500, and the newer R1 (named one of TIME's Best Inventions of 2025) is available for just $4,900–$5,900. Unitree plans to ship 10,000–20,000 humanoid robots in 2026 (up from ~5,500 in 2025) and has open-sourced UnifoLM-VLA-0, a Vision-Language-Action model for general-purpose manipulation.

The humanoid robotics field has exploded: CES 2026 listed 38 companies in the humanoid robotics category, with China controlling approximately 90% of the global market. Key Chinese entrants beyond Unitree include Agibot, Galbot, Engine AI, and Fourier, while European companies like NEURA Robotics (Germany) are entering with robots priced from EUR 19,999.

Key Takeaways

• Waymo is the global leader in commercial autonomous vehicles — 170 million+ fully autonomous miles across 10 US cities, $126 billion valuation, 91% fewer serious-injury crashes than human drivers, with international expansion to London planned
• Tesla FSD uses a vision-only, data-scale approach; Cybercab volume production begins April 2026. Zoox has launched commercially in Las Vegas and SF with 300K+ riders and an Uber partnership
• Aurora Innovation has logged 250K+ incident-free driverless miles in autonomous trucking, targeting 200+ trucks by end of 2026
• Humanoid robotics has entered its commercial phase: Figure AI ($39 billion valuation, Figure 03 for home use at $20K), Boston Dynamics Atlas (now in production, shipping to Hyundai and Google DeepMind), Tesla Optimus (Gen 3 in production), Unitree ($13,500 G1, 10-20K units planned in 2026), and 38 companies at CES 2026 with China controlling ~90% of the market
• Physical AI faces a fundamental data challenge — physical interaction data is expensive to collect compared to the text data that trained LLMs, making robotics harder to scale than language AI