Humanoid Robots — Where Are We Really?
Separate the hype from the reality. Where does humanoid robotics actually stand in 2025, and what are the remaining hard problems?
The Dream and the Reality
The humanoid robot has been the defining image of science fiction AI for 80 years. C-3PO. Data. The Terminator. A machine with a human body, walking, talking, doing human tasks in a human world.
In 2025, humanoid robots are real. They exist, they walk, some can manipulate objects. Billions of dollars are being invested by serious companies.
They are also, in many ways, still very much a work in progress.
This lesson separates what’s working from what’s not — and why the remaining problems are so hard.
The Appeal of Human Form
Why build a robot that looks like a human? Three reasons:
1. The world is designed for humans. Doorknobs, stairs, ladders, tools, vehicles, kitchens — virtually every human environment and tool was designed for a human body. A humanoid robot can operate in these environments without redesigning them.
2. Human tasks are diverse. A robot built for a specific task is efficient but inflexible. A humanoid robot is, in theory, a general-purpose worker that can do any physical task a human can do.
3. Human interaction. Humans are social animals who naturally interact with things that look like them. A humanoid robot may be more intuitive to communicate with and collaborate alongside than an abstract arm or wheeled platform.
What’s Working
Bipedal locomotion Boston Dynamics’ Atlas can run, jump, do backflips, and navigate rough terrain. Figure, 1X, Agility Robotics, and Unitree have all demonstrated stable bipedal walking in real-world environments. Walking on flat floors, climbing stairs, and recovering from pushes are all solved or nearly solved.
Basic manipulation Humanoid robots can pick up boxes, stack items, open doors, and carry objects. These are not trivial — they represent years of progress in grasping, balance adjustment while holding loads, and coordinated whole-body motion.
LLM integration Large language models are being integrated as the “brain” for humanoid robots — allowing them to understand natural language instructions and reason about tasks. You can tell some robots what to do in plain English.
What’s Still Hard
Dexterous manipulation Hands are extremely difficult. A human hand has 27 degrees of freedom and 17,000 touch receptors per fingertip. Current humanoid robot hands are dramatically less capable: fewer fingers, less dexterity, almost no tactile sensing.
Tasks like using a key, buttoning a shirt, or unscrewing a jar cap remain very difficult for current humanoid robots. These require coordination, force modulation, and tactile feedback that is far beyond current hardware.
Speed and energy efficiency Current humanoid robots walk slowly — typically 1-1.5 m/s, compared to 1.4 m/s average human walking speed. Running and fast manipulation are either impossible or very limited. Battery life is typically 1-2 hours of active operation.
Generalization This is the biggest challenge. A humanoid robot trained to stack boxes in a factory can’t spontaneously figure out how to fold laundry. Each new task requires significant retraining or programming. The vision of a robot that can do anything a human can do — transferring knowledge across domains fluidly — remains far from reality.
Stanford researchers propose a simple test for humanoid robot capability: can the robot put a shirt on a hanger?
It requires:
- Identifying the shirt and hanger despite deformable shape
- Grasping a floppy, non-rigid object
- Threading the hanger through the collar (precise dexterous manipulation)
- Adjusting grip as the shirt moves (tactile feedback)
- Completing the task in reasonable time
No current humanoid robot can do this reliably. It’s a useful benchmark for tracking real progress against the hype.
The Key Players (2025)
Boston Dynamics (Atlas) The most technically impressive locomotion — Atlas can run, jump, and handle rough terrain. Focus on dynamic movement and whole-body control. Hyundai-owned.
Figure AI BMW partnership for automotive factory deployment. One of the best-funded startups. Integrating OpenAI models for language understanding.
Agility Robotics (Digit) Amazon partnership for warehouse logistics. Focused on a narrow use case (moving totes in fulfillment centers) rather than general purpose. More commercially realistic near-term.
1X Technologies Norwegian company with backing from OpenAI. Focus on “androidal” robot for general home and industrial use.
Tesla Optimus Tesla’s humanoid project. Slower to show capability than some competitors, but Tesla’s manufacturing scale and AI investment are significant long-term advantages.
Unitree (H1, G1) Chinese manufacturer producing the lowest-cost humanoid robots. H1 costs under $100,000, making humanoid hardware accessible for research.
Robot demos are almost universally misleading.
A demo video shows a robot doing an impressive task smoothly and successfully. What it doesn’t show:
- How many takes were required
- How controlled the environment was
- Whether the robot can do this reliably (99%+ of the time) or occasionally (10-30%)
- What happens when anything goes slightly wrong
The gap between “can do this in ideal conditions occasionally” and “can do this reliably in real-world conditions” is enormous — often years of engineering work.
When evaluating humanoid robot progress, ask not “can it do this?” but “what is the success rate, under what conditions, with how much human intervention?”
Why is dexterous manipulation considered one of the hardest remaining problems in humanoid robotics?
Reveal Answer & Explanation
Dexterous manipulation requires simultaneous capabilities that current hardware and AI struggle to provide: many degrees of freedom in the hand (27 in the human hand), fine tactile sensing (17,000 touch receptors per human fingertip), real-time force modulation (adjusting grip as objects slip or shift), and the ability to handle deformable and unpredictable objects.
Most humanoid robot hands have far fewer joints, minimal tactile sensing, and struggle with the force control required for delicate manipulation. Tasks as simple as using a key, buttoning a shirt, or handling a raw egg remain beyond reliable capability for current systems.
What’s Next
Module 4 covered the robots you build. In Module 5, we focus on robots that navigate — specifically, autonomous vehicles that have to find their way through the most complex environment of all: the real world.