Humanoid Robots
Two-legged robots with AI control enter first volume production, but so far only work narrowly defined factory jobs.
Open in the interactive tree →Since 2024/25 companies such as Figure, Agility, Tesla, Unitree and AgiBot have promised humanoid robots as general-purpose workers; AI models steer them from language and camera images. In practice they do transport and handling jobs in warehouses and plants. Price and reliability will decide whether this becomes a mass market.
As of October 2026
Unitree shipped more than 5,500 humanoids in 2025 (G1 from around $16,000) and listed on Shanghai’s STAR Market on 19 August 2026. Press reports citing market estimates (originating research firm unclear) put first-half 2026 shipments at about 19,100 worldwide, over 97% from Chinese makers (AgiBot about 8,400, Unitree about 5,900). According to the companies themselves, Figure’s robots logged over 1,250 hours at a BMW plant with above 99% placement accuracy, and Agility’s Digit has passed 65,000 operating hours across nine customer sites. Tesla’s Optimus was “not in usage in our factories in a material way” per Musk, and no verified production unit had been shown by Tesla’s 22 July 2026 earnings call.
Open steps
- Training data for manipulation High AI leverageRobots learn from teleoperation, human video and simulation; none gives enough varied, high-quality experience for general tasks at low cost.
- Closing the gap from simulation High AI leveragePolicies trained in simulation fail on real friction, compliance and wear; automatic tuning of simulators and rewards must transfer reliably.
- Hands that last a million grasps Medium AI leverageDexterous hands with touch sensing wear out and cost too much; tendon, actuator and skin designs must survive years of factory work.
- Safety tests for balancing robots Low AI leverageA balancing robot falls when power or control fails; the draft ISO standard for such robots is unfinished, so safe use near people lacks agreed tests.
Where AI could help
Medium AI leverage. Learned policies, simulation and synthetic data drive capability; actuators, batteries, safety and unit cost set deployment.
- Train whole-body control and manipulation in simulation, then transfer
- Generate synthetic and video-model data for rare situations
- Learn from fleet experience and human corrections
- Automate safety testing and fault diagnosis
Shown so far
- In March 2025 NVIDIA's GR00T N1 humanoid foundation model was trained on a mix of real-robot trajectories, human videos and synthetic data (preprint, according to the company). source
- In November 2025 Physical Intelligence reported (company paper) that learning from its own experience and corrections more than doubled throughput and roughly halved failures on hard tasks such as laundry folding and espresso. source
- In June 2024 DrEureka (RSS 2024) used an LLM to design rewards and domain randomization for sim-to-real transfer, finding configurations competitive with human-designed ones on real quadruped tasks. source