Last reviewed
Jun 7, 2026
Sources
30 citations
Review status
Source-backed
Revision
v3 ยท 7,317 words
Add missing citations, update stale details, or suggest a clearer explanation.
Humanoid robot autonomy levels describe a graduated framework for classifying the degree of independent operation that a humanoid robot can achieve. Ranging from Level 0 (no autonomy, full human control) through Level 5 (full general intelligence), the framework provides a standardized way to compare different robots and track the progress of the industry over time. As of 2025, most commercially available humanoid robots operate at Level 2 or Level 3, with several companies actively developing Level 4 systems. Beyond autonomy, humanoid robots can also be classified by their physical form, application domain, and the nature of their artificial intelligence integration.
The autonomy framework for humanoid robots draws loose parallels from the SAE levels of driving automation used in the automotive industry, adapted for the broader and more complex domain of general-purpose robotics. Each level represents a qualitative increase in the robot's ability to perceive, reason, and act without human intervention.
At Level 0, the robot has no capacity for independent operation. Every movement and action requires continuous, real-time human control. The operator must directly command each joint position or end-effector trajectory. Robots at this level are essentially sophisticated puppets, useful for teleoperation research and performance art but incapable of any autonomous task execution. Early humanoid prototypes from the 1960s and 1970s fell into this category, as did many research platforms built purely for mechanical validation.
Level 1 robots possess basic programmable movement with limited functionality. They can execute pre-programmed motion sequences (such as walking a fixed path or waving) but lack the ability to adapt to changes in their environment. Their control systems typically rely on open-loop or simple closed-loop algorithms with minimal sensor feedback. Industrial robotic arms from the 1980s and 1990s operated at roughly this level of capability. In the humanoid domain, early versions of Honda ASIMO (circa 2000) represented a transition from L0 to L1, executing choreographed walking patterns on flat, predictable surfaces.
Level 2 robots use algorithm-driven movement within structured environments. They incorporate sensor feedback loops for balance correction and obstacle detection but still operate within tightly constrained parameters. These robots can perform specific tasks in controlled settings (factory floors, laboratories) with minimal human oversight during execution, though a human must set up the task and monitor for failures. SoftBank Robotics NAO, widely used in academic research and education, operates near this level, capable of navigating known indoor environments and performing scripted interactions. Many warehouse and logistics robots with partial humanoid features also fall into this category.
Level 3 represents the current industry standard for leading humanoid robots as of 2025. Robots at this level are equipped with sensor suites (cameras, LiDAR, force/torque sensors, IMUs) that provide real environmental awareness. They can adapt their behavior based on perceived conditions, handle moderate variability in their surroundings, and recover from some classes of errors without human intervention. However, they still require human oversight for novel situations, complex decision-making, and safety-critical judgments.
Examples of Level 3 systems include early deployment versions of Tesla Optimus [5], the Boston Dynamics Atlas platform in its later iterations [6], and initial versions of Agility Robotics Digit [11]. These robots can perform tasks like picking and placing objects, navigating semi-structured environments, and responding to basic voice commands. Their autonomy is conditional because it depends on operating within a known set of scenarios; truly unexpected situations still require human takeover or cause the robot to halt.
Level 4 robots can complete tasks with independent reasoning, planning, and execution across a broader range of environments. They can handle novel objects, recover from a wider set of failures, and make contextual decisions about how to accomplish goals that were specified at a high level (for example, "clean the kitchen" rather than "pick up object A and place it in location B"). This level requires sophisticated world models, advanced manipulation skills, and robust perception systems that can handle occlusion, clutter, and dynamic changes.
As of 2025, Level 4 is actively under development by several companies. Figure 02 [4] with its OpenAI-powered reasoning capabilities, advanced versions of Tesla Optimus with end-to-end neural network control, and Sanctuary AI Phoenix [12] with its Carbon AI system all aim to reach this tier. No humanoid robot has yet achieved consistent, reliable Level 4 operation across diverse real-world environments, though demonstrations in controlled settings have shown promising results.
Level 5 represents theoretical human-equivalent general intelligence in a robotic body. A Level 5 humanoid robot would possess creative problem-solving abilities, abstract reasoning, emotional understanding, and the capacity to learn any new task from minimal instruction, just as a human can. It would seamlessly handle novel environments, social interactions, and multi-step tasks that require common sense reasoning.
No current technology approaches Level 5. Achieving it would require breakthroughs in artificial general intelligence (AGI), and most researchers consider it decades away at minimum. The gap between Level 4 and Level 5 is arguably the largest in the entire framework, as it encompasses the full breadth of human cognitive capability.
| Level | Name | Description | Key capabilities | Examples | Status (2025) |
|---|---|---|---|---|---|
| L0 | No autonomy | Continuous human control required for all movements and actions | Teleoperation only; no independent sensing or decision-making | Early research prototypes, teleoperated humanoids | Historical |
| L1 | Auxiliary control | Basic programmable movement with limited function | Pre-programmed motion sequences; minimal sensor feedback; open-loop or simple closed-loop control | Early Honda ASIMO, basic research platforms | Historical/Legacy |
| L2 | Partial autonomy | Algorithm-driven movement within structured environments | Sensor-based balance correction; obstacle detection in known settings; task execution with human setup and monitoring | SoftBank Robotics NAO, early warehouse humanoids, Fourier Intelligence GR-1 | Available |
| L3 | Conditional autonomy | Sensor-equipped with environmental awareness and adaptive behavior | Environmental perception via cameras/LiDAR/IMU; moderate error recovery; voice command response; task adaptation within known scenarios | Tesla Optimus (early deployment), Boston Dynamics Atlas, Agility Robotics Digit | Current industry standard |
| L4 | High autonomy | Independent reasoning and task completion across diverse environments | High-level goal interpretation; novel object handling; multi-step planning; broad failure recovery; contextual decision-making | Figure 02 (target), advanced Tesla Optimus (target), Sanctuary AI Phoenix (target) | In development |
| L5 | Full intelligence | Human-equivalent general intelligence with creative problem-solving | Abstract reasoning; emotional understanding; learning any task from minimal instruction; common sense reasoning; full social interaction | None | Theoretical |
The autonomy framework described above is one of several attempts to grade humanoid intelligence. In May 2025 the China Electronics Society published T/CIE 298-2025, described as the world's first humanoid robot intelligence grading standard, led by the Beijing X-Humanoid (Beijing Humanoid Robot Innovation Center) together with innovation centers in Shanghai and Zhejiang and companies including UBTECH and Unitree. [16] The standard adopts a "four-dimension, five-level" model: it evaluates robots across four ability dimensions, perception and cognition, decision-making and learning, execution and performance, and collaboration and interaction, and assigns an overall L1 to L5 intelligence grade. [16] It specifies 22 primary indicators, more than 100 technical provisions, and a universal safety baseline, and like the autonomy framework it explicitly borrows grading logic from autonomous driving and industrial robotics while accounting for the distinct demands of machines that walk, grasp, and share space with people. [16]
Beyond autonomy levels, humanoid robots can be classified by their physical locomotion design. The two primary categories are bipedal and wheeled humanoids, each with distinct advantages and trade-offs.
Bipedal humanoid robots walk on two legs, closely mimicking the human gait. This design offers maximum mobility and the ability to navigate environments built for humans, including stairs, narrow doorways, uneven terrain, and cluttered workspaces. However, bipedal locomotion imposes complex balance requirements. Maintaining stability while walking, turning, carrying loads, or recovering from pushes demands sophisticated control algorithms and high-bandwidth actuators.
Most flagship humanoid robots adopt the bipedal form factor. Boston Dynamics Atlas, Tesla Optimus, Figure 02, Unitree H1, and UBTECH Walker S2 are all bipedal. The engineering challenge of bipedal walking has driven significant advances in reinforcement learning, model predictive control, and actuator design. The Unitree H1 holds a confirmed world speed record for bipedal humanoid running of 3.3 meters per second. [30]
Wheeled humanoid robots feature a human-like upper body mounted on a wheeled base. This design sacrifices stair navigation and rough terrain capability in exchange for greater stability, simpler control, lower energy consumption, and smoother movement on flat surfaces. Wheeled humanoids are particularly well-suited for structured indoor environments such as offices, hospitals, retail stores, and airports.
Examples include Enchanted Tools Mirokai, certain configurations of service robots from Keenon Robotics, and the LG Electronics CLOi platform. For many commercial service applications, the wheeled form factor offers a practical compromise between human-like interaction capability and operational reliability.
| Form factor | Advantages | Disadvantages | Best suited for | Examples |
|---|---|---|---|---|
| Bipedal | Navigates stairs and uneven terrain; fits in human-designed spaces; maximum versatility | Complex balance control; higher energy consumption; risk of falls; more expensive actuators | Manufacturing, home assistance, construction, disaster response | Tesla Optimus, Figure 02, Unitree H1 |
| Wheeled | Greater stability; simpler control; lower energy use; smoother movement; lower cost | Cannot climb stairs; limited on uneven terrain; restricted to flat indoor environments | Retail, hospitality, healthcare, offices, airports | Enchanted Tools Mirokai, LG Electronics CLOi |
Humanoid robots are designed for different deployment environments, each placing distinct demands on the robot's motion control, perception, and autonomy capabilities. These domains can be ranked by the complexity of the motion control required.
Industrial manufacturing environments place the lowest relative motion demands on humanoid robots because the workspace is highly structured, predictable, and engineered for efficiency. Tasks include loading parts onto assembly lines, operating machinery, moving totes between stations, and performing quality inspections. The environment is typically well-lit, temperature-controlled, and free of unexpected obstacles. Figure 02 deployed at BMW's Spartanburg plant [4] and Agility Robotics Digit working in Amazon fulfillment centers [11] represent leading examples in this domain.
Commercial service applications require robots to operate in semi-structured settings with moderate human traffic, such as retail stores, hotels, hospitals, and restaurants. Robots must navigate around people, interact socially, and handle a wider variety of objects than in a factory. The environment is less predictable than a factory floor but still follows regular patterns. UBTECH Walker and various service-oriented platforms target this domain.
Extreme environment applications deploy humanoid robots in hazardous locations where human presence is dangerous or impossible. This includes nuclear facilities, disaster sites, deep-sea operations, space exploration, and firefighting scenarios. These environments demand robust hardware (radiation hardening, waterproofing, temperature resistance) and high autonomy, since communication delays or signal loss may prevent real-time human control. NASA Valkyrie (R5) was designed specifically for disaster response and space exploration, while Boston Dynamics Atlas has been tested in simulated disaster scenarios through DARPA challenges. [2]
Home service represents the highest motion control requirements because domestic environments are inherently unstructured, cluttered, and variable. A home robot must handle deformable objects (clothing, food), navigate around furniture and pets, operate in varying lighting conditions, and interact safely with children and elderly individuals. [1] No humanoid robot has achieved reliable autonomous operation in home environments as of 2025, though this remains one of the most active areas of research and a primary long-term market target for companies like Tesla, 1X Technologies, and Figure.
| Domain | Environment structure | Motion complexity | Key challenges | Leading examples |
|---|---|---|---|---|
| Industrial manufacturing | Highly structured | Lowest | Repetitive precision, cycle time, integration with existing lines | Figure 02 (BMW), Agility Robotics Digit (Amazon) |
| Commercial service | Semi-structured | Moderate | Human navigation, social interaction, object variety | UBTECH Walker S2, service platforms |
| Extreme environments | Unstructured/hazardous | High | Robustness, communication latency, hardware durability | NASA Valkyrie R5, Boston Dynamics Atlas |
| Home services | Unstructured/variable | Highest | Deformable objects, clutter, safety with people, diverse tasks | Research stage; targeted by Tesla, 1X Technologies, Figure |
The humanoid robotics industry features three distinct business models for AI integration, reflecting different strategic bets on where value will accumulate in the robotics stack.
Vertically integrated companies build both the robot hardware and the AI software that controls it. This approach allows tight co-optimization of the mechanical design, sensor suite, and learning algorithms. Tesla (Optimus) [5], Figure (Figure 02/03) [4], and Unitree all pursue this strategy, developing proprietary neural networks trained on data collected from their own hardware. The advantage is end-to-end control over performance and rapid iteration cycles. The disadvantage is the enormous capital and talent required to excel simultaneously in mechanical engineering, electrical engineering, and AI research. In February 2025 Figure made this bet explicit, ending its collaboration with OpenAI to develop AI in-house; chief executive Brett Adcock argued that solving embodied AI at scale "in the real world" required vertically integrating robot AI. [18]
Hardware-focused companies specialize in building high-quality robot bodies and rely on partnerships or third-party solutions for higher-level AI capabilities. Sanctuary AI develops its Carbon AI system but places primary emphasis on its dexterous hardware. [12] 1X Technologies builds the NEO platform with a focus on safe, lightweight mechanical design. Boston Dynamics has decades of expertise in dynamic locomotion hardware and increasingly integrates AI through collaborations with partners. These companies bet that the hardware challenge is the primary bottleneck and that AI capabilities can be sourced or developed incrementally. In 2025 Sanctuary AI presented the eighth generation of Phoenix and integrated tactile sensors aimed at speeding up teleoperated data collection, reflecting a strategy that leans on remote piloting and touch sensing to gather the demonstrations its autonomous models learn from. [29]
AI model providers develop the foundation models and software platforms that power humanoid robots but do not build the physical hardware themselves. OpenAI provides language and reasoning models to Figure through a partnership. NVIDIA offers the Isaac and GR00T platforms for robot simulation, training, and deployment. [10] Google DeepMind develops robotics foundation models like RT-2 and RT-X. [3] Microsoft invests in robotics AI through Azure cloud services and partnerships. These companies view humanoid robots as a deployment platform for their AI capabilities, similar to how smartphone operating systems run on hardware from multiple manufacturers.
| Integration model | Strategy | Advantages | Disadvantages | Key players |
|---|---|---|---|---|
| Vertically integrated | Build both hardware and AI | End-to-end optimization; rapid iteration; proprietary data flywheel | Enormous capital requirements; must excel across multiple disciplines | Tesla, Figure, Unitree |
| Hardware-focused | Specialize in robot bodies; partner for AI | Deep mechanical expertise; can adopt best-available AI; lower AI R&D costs | Dependent on external AI partners; less control over software performance | Sanctuary AI, 1X Technologies, Boston Dynamics |
| AI model providers | Develop foundation models and platforms | Scalable across many hardware platforms; leverage existing AI research; recurring software revenue | No control over hardware quality; dependent on hardware partners for deployment | OpenAI, NVIDIA, Google DeepMind, Microsoft |
Note on Figure and OpenAI: the partnership described above reflected the situation through early 2025. In February 2025 Figure ended the collaboration to build its own vision-language-action model, Helix, so OpenAI no longer supplies the models that run on Figure's robots. [18] See the 2025-2026 developments section below.
The AI systems powering humanoid robots employ several distinct learning paradigms, each suited to different aspects of robot behavior. Modern systems increasingly combine multiple approaches.
Reinforcement learning (RL) enables robots to improve through trial and error. The robot takes actions in an environment, receives reward signals based on outcomes, and gradually learns policies that maximize cumulative reward. In humanoid robotics, RL has proven particularly effective for locomotion (walking, running, balancing, stair climbing) and dynamic tasks where the optimal control strategy is difficult to specify analytically.
RL training typically occurs in physics simulation environments before being transferred to the physical robot, a process known as sim-to-real transfer. NVIDIA Isaac Sim and MuJoCo are commonly used simulation platforms. The challenge lies in bridging the "reality gap" between simulated and real-world physics. Domain randomization (varying simulation parameters during training) and system identification (calibrating the simulator to match the real robot) are standard techniques for addressing this gap.
Imitation learning (also called learning from demonstration) allows robots to acquire skills by observing human demonstrations. Rather than discovering optimal behaviors through trial and error, the robot directly learns a mapping from observations to actions based on expert examples. This approach is especially useful for manipulation tasks where specifying a reward function is difficult but demonstrating the desired behavior is straightforward.
Common methods include behavioral cloning (supervised learning on demonstration data) and inverse reinforcement learning (inferring the reward function from demonstrations). Tesla has used human teleoperators wearing motion-capture suits to collect demonstration data for Optimus. [5] Figure has employed VR teleoperation to gather training data for dexterous manipulation. [4] The scalability of data collection is a key bottleneck, and recent work on using internet video data for imitation learning aims to address this limitation.
Vision-language models (VLMs) enable robots to understand natural language commands and reason about the visual world. These models, built on large language model architectures extended with visual encoders, can interpret instructions like "pick up the red cup on the table" by jointly processing camera images and text input. VLMs serve as the cognitive layer that bridges high-level human intent and low-level robot actions.
In humanoid robotics, VLMs are used for task planning, scene understanding, and human-robot dialogue. OpenAI's models power Figure 02's conversational abilities. Google DeepMind's PaLM-E and RT-2 demonstrated that VLMs can directly output robot control commands. [3] The integration of VLMs gives humanoid robots a form of common-sense reasoning, allowing them to infer context, handle ambiguous instructions, and generalize to novel situations not seen during training.
Vision-language-action (VLA) models extend VLMs by directly outputting robot actions from visual and language inputs in an end-to-end architecture. This eliminates the need for hand-crafted perception pipelines, separate planning modules, and explicit state estimation. The model takes in camera images and a language instruction, then directly produces motor commands or action tokens.
Google DeepMind's RT-2 (Robotic Transformer 2) was a landmark VLA model, demonstrating that a single model could reason about the world and produce physical actions. [3] Subsequent work, including RT-X (trained on data from multiple robot embodiments) and Physical Intelligence's pi0 model, has pushed VLA capabilities further. VLA models represent a shift toward unified, general-purpose robot brains that can potentially transfer across different robot bodies and tasks.
Large Behavior Models (LBMs) represent the newest frontier in humanoid robot AI. Pioneered by Boston Dynamics in collaboration with the Toyota Research Institute, LBMs provide unified control of the entire robot body, treating hands and feet as parts of an integrated whole rather than as separate subsystems. [13] This approach allows the robot to coordinate complex, multi-contact behaviors, such as using one hand for support while the other manipulates an object, or transitioning smoothly between walking and reaching.
LBMs are trained on large-scale datasets of whole-body behavior, capturing the continuous, fluid nature of human movement rather than decomposing it into discrete skills. The key advantage is the ability to execute long, complex task sequences without the brittle handoffs between separate locomotion and manipulation controllers that plague traditional approaches. [13] Toyota Research Institute has demonstrated LBM-controlled robots performing extended kitchen tasks that require simultaneous walking, reaching, grasping, and placing. On August 20, 2025, Boston Dynamics and TRI showed an electric Atlas controlled by a single LBM performing a continuous sequence of packing, sorting, and organizing tasks that combine manipulation with walking, crouching, and lifting, and recovering autonomously when researchers introduced unexpected disturbances mid-task. [17] Russ Tedrake, senior vice president of large behavior models at TRI, said skills are added through human demonstrations and that "as the LBMs get stronger, they require less and less demonstrations." [17]
| Learning method | How it works | Strengths | Limitations | Key applications |
|---|---|---|---|---|
| Reinforcement learning | Trial-and-error optimization via reward signals in simulation | Discovers novel solutions; excels at dynamic tasks; handles complex physics | Reward design is difficult; sim-to-real gap; sample inefficient | Walking, running, balancing, stair climbing, dynamic recovery |
| Imitation learning | Supervised learning from human demonstrations | Intuitive data collection; works where reward functions are hard to define; fast initial learning | Limited by demonstration quality and quantity; distribution shift problems | Dexterous manipulation, tool use, household tasks |
| Vision-language models | Joint processing of visual and textual inputs for reasoning | Common-sense reasoning; handles ambiguous instructions; generalizes to novel contexts | Computationally expensive; may hallucinate; slow inference | Task planning, scene understanding, human-robot dialogue |
| Vision-language-action models | End-to-end mapping from images and language to motor commands | Eliminates hand-crafted pipelines; transfers across embodiments; unified architecture | Requires massive training data; challenging to ensure safety; early stage | General-purpose manipulation, multi-step task execution |
| Large behavior models | Unified whole-body control from large-scale behavior datasets | Integrated locomotion and manipulation; fluid, continuous movement; long task sequences | Data-intensive; early research stage; limited availability | Complex multi-contact tasks, kitchen work, industrial assembly |
The low-level control systems that translate high-level AI decisions into physical joint movements are critical to humanoid robot performance. Several control paradigms are used, each with distinct characteristics.
Impedance control makes the robot behave like a mass-spring-damper system, regulating the relationship between the robot's position and the force it exerts on the environment. Rather than commanding a rigid position trajectory, the controller specifies a desired mechanical impedance (stiffness and damping) that determines how the robot responds to external forces. This makes the robot inherently compliant and safe for physical interaction.
Impedance control is widely used in humanoid robots that must interact with people or handle delicate objects. It provides stable behavior under unexpected contacts, such as being bumped or encountering an unseen obstacle. [7] The primary drawback is that it requires force/torque sensing at the joints or end-effectors, which adds cost, complexity, and potential failure points to the hardware.
Admittance control takes the inverse approach: it measures external forces (via force/torque sensors) and converts them into reference velocities or positions for the robot's position-controlled joints. The robot "admits" external forces by moving in the direction of the applied force, creating compliant behavior on top of stiff position-controlled actuators.
This approach is particularly useful for robots that use high-gain position-controlled servos, which are common in industrial humanoid platforms. The advantage is compatibility with standard position-controlled hardware. The disadvantage is that the behavior can feel sluggish or delayed compared to true impedance control, since the compliance is computed rather than inherent in the mechanical system. Tuning the admittance parameters for natural-feeling interaction remains an active area of research.
Whole-Body Quadratic Programming (QP) Control formulates the robot's control problem as a mathematical optimization solved at each control timestep. The QP balances multiple objectives simultaneously: maintaining balance, following a desired trajectory, respecting joint limits, obeying contact friction constraints, and minimizing energy consumption. Each objective is expressed as a cost function or constraint, and the QP solver finds the joint torques that best satisfy all of them. [14]
This approach is the backbone of modern dynamic humanoid locomotion. Boston Dynamics Atlas and many research humanoids use variants of whole-body QP control. [15] The strength is the ability to handle complex, multi-task scenarios elegantly within a unified framework. The weakness is computational intensity; solving a QP at 500 Hz or higher requires powerful onboard processors, and the formulation must be carefully constructed to remain feasible under all conditions.
Human-aware control extends standard controllers to incorporate models of human state and dynamics. The robot considers not just its own objectives but also the posture, fatigue level, intent, and comfort of nearby humans. This is essential for collaborative humanoid robots working alongside people in manufacturing or home environments.
The controller may reduce speed and force when a human is nearby, adjust its posture to maintain clear sightlines, or modify its grasp strategy to facilitate handovers. [7] Advanced versions model human ergonomics and predict fatigue, adjusting the collaboration pattern to reduce strain on the human partner. The primary challenge is accurate real-time estimation of human state, which requires sophisticated perception systems and predictive models.
| Control approach | Mechanism | Pros | Cons | Typical use cases |
|---|---|---|---|---|
| Impedance control | Regulates position-force relationship as a virtual mass-spring-damper | Stable under unexpected contacts; inherently compliant; safe for human interaction | Requires force/torque sensing hardware; complex parameter tuning; sensor cost and fragility | Physical human-robot interaction, delicate object handling, collaborative tasks |
| Admittance control | Converts measured forces into reference velocities for position-controlled joints | Compatible with standard position-controlled hardware; straightforward implementation | Can feel sluggish or delayed; less natural compliance than impedance control; tuning challenges | Industrial robots with position-controlled servos, assembly tasks |
| Whole-body QP control | Solves quadratic program at each timestep to balance multiple objectives | Handles complex multi-task scenarios; unified framework; respects physical constraints | Computationally intensive; requires powerful onboard processors; careful formulation needed | Dynamic locomotion, whole-body coordination, parkour-style movements |
| Human-aware control | Incorporates human state and ergonomics into control objectives | Considers human comfort and safety; reduces partner fatigue; enables natural collaboration | Requires accurate real-time human state estimation; complex perception systems; limited validation | Human-robot collaboration in manufacturing, assistive robotics, home environments |
As humanoid robots increasingly operate alongside people, the design of effective human-robot interaction (HRI) has become a major research and engineering focus. [8] HRI for humanoid robots spans three primary domains, each with distinct interaction paradigms.
Companions. Companion humanoid robots serve in coaching, therapy, elderly care, and educational roles. They engage in long-term social relationships with users, requiring consistent personality, emotional responsiveness, and the ability to remember previous interactions. Applications include rehabilitation coaching for stroke patients, social engagement for individuals with autism, and companionship for isolated elderly populations. The robot's social skills are paramount; it must convey warmth, patience, and attentiveness through facial expressions, gestures, and speech.
Co-Workers. Co-worker humanoid robots collaborate with humans in manufacturing, logistics, and professional settings. The interaction is task-focused, emphasizing efficiency, safety, and clear communication of intent. The robot must signal its planned movements to nearby workers, respond to human gestures and commands, and adjust its pace to match the human partner. Figure 02's deployment at BMW, where robots work alongside human operators on the assembly line, exemplifies this domain. [4]
Avatars. Avatar humanoid robots are teleoperated platforms that provide a physical presence for a remote human operator. The operator sees through the robot's cameras, speaks through its speakers, and controls its movements in real time. Applications include remote medical consultations, hazardous environment inspection, and virtual tourism. Toyota T-HR3 was designed specifically as a telepresence avatar, allowing operators to control the robot's movements through a master-slave system with haptic feedback.
Effective humanoid robots require several social and cognitive capabilities that go beyond task execution.
Believability refers to the robot's ability to exhibit consistent, coherent behaviors that make it seem like a coherent agent rather than a collection of disconnected responses. A believable robot maintains a stable personality, remembers context from earlier in a conversation, and behaves in ways that align with its established character. Inconsistent behavior erodes human trust and engagement.
Readability is the capacity to signal intentions clearly so that nearby humans can predict the robot's next actions. A readable robot telegraphs its movements through gaze direction, preparatory gestures, and explicit verbal cues. In collaborative manufacturing, readability is a safety requirement; workers must be able to anticipate the robot's movements to avoid collisions.
Theory of Mind (ToM) is the ability to attribute mental states (beliefs, desires, intentions) to other agents. A robot with ToM can infer that a human partner is confused, frustrated, or waiting for assistance, and adjust its behavior accordingly. This capability is essential for natural collaboration and is one of the most challenging aspects of humanoid robot AI. Current systems approximate ToM through learned models of human behavior rather than true mental state reasoning.
The development of humanoid robots can be understood through a six-stage evolutionary framework that describes the progressive accumulation of capabilities.
Stage 1: Structures. The foundation of humanoid robotics is mechanical design. This stage focuses on building bodies with human-like proportions, degrees of freedom, and physical capabilities. Advances in materials science, actuator technology, and mechanical engineering drive progress at this stage. The transition from hydraulic to electric actuation and from rigid to compliant mechanisms represents major milestones.
Stage 2: Senses. The second stage adds perception. Robots gain the ability to see (cameras, depth sensors), feel (force/torque sensors, tactile arrays), hear (microphones), and sense their own body state (IMUs, joint encoders). The richness and reliability of sensory input determine the upper bound on what the robot can achieve autonomously.
Stage 3: Behaviors. With structure and senses in place, robots develop repertoires of behaviors: walking, grasping, balancing, reaching, and manipulating. These behaviors may be hand-crafted (traditional robotics) or learned (reinforcement learning, imitation learning). The transition from scripted to adaptive behaviors marks a critical inflection point.
Stage 4: Functions. Behaviors combine into useful functions: performing a complete task like loading boxes, cooking a meal, or assisting a patient. This stage requires task planning, sequencing, and error recovery. Most commercial humanoid robots in 2025 are working at the boundary between Stage 3 and Stage 4.
Stage 5: Humanity. At this stage, robots develop social and emotional capabilities that enable rich interaction with humans. They display empathy, humor, cultural awareness, and the ability to form ongoing relationships. This stage goes beyond functional task completion to encompass the full range of human social behavior.
Stage 6: Intelligence. The final stage represents general intelligence, the ability to learn, reason, create, and adapt across any domain without specific programming. This corresponds to Level 5 autonomy and remains a theoretical goal.
The evolution framework also defines three paradigm levels that describe the overall character of a humanoid robot:
| Paradigm level | Description | Key characteristics | Current status |
|---|---|---|---|
| Human-Looking | Robot has a human-like physical appearance | Correct proportions, recognizable body parts, human-like face and hands | Achieved by many platforms |
| Human-Like | Robot exhibits human-like behaviors and capabilities | Natural movement, social interaction, adaptive task execution, emotional expression | Partially achieved; active development |
| Human-Level | Robot matches human general intelligence and versatility | Creative problem-solving, abstract reasoning, learning from minimal instruction, full social competence | Theoretical; equivalent to AGI |
As the humanoid robotics industry generates increasing numbers of impressive-looking demonstration videos and announcements, a critical evaluation framework helps distinguish genuine progress from marketing. [9] The following checklist provides a structured approach to assessing robot demonstrations.
Actual deployment location. Where was the demo conducted? A controlled lab with perfect lighting, flat floors, and no obstacles is very different from a real factory, warehouse, or home. Demos on company premises with engineered conditions should be viewed with appropriate skepticism about real-world generalization.
Who controlled the demo. Was the robot operating autonomously, or was it teleoperated (partially or fully) by a human? Many impressive demonstrations involve hidden human operators or extensive pre-programming. Look for explicit statements about the level of autonomy during the shown tasks.
Failure documentation. Does the company show failures alongside successes? All robots fail, and transparency about failure modes indicates engineering maturity. Companies that only show perfect runs may be cherry-picking from many attempts.
Real product offering. Is the robot available for purchase or lease, or is it purely a research prototype? A functioning product with paying customers represents a fundamentally different stage of development than a lab demonstration.
Data requirements. How much training data does the robot need to learn a new task? A system that requires thousands of demonstrations for each new skill has very different scalability prospects than one that generalizes from a handful of examples.
Fine print and restrictions. What constraints apply to the demo conditions? Look for disclaimers about controlled environments, pre-mapped spaces, limited object sets, or specific lighting requirements. The gap between demo conditions and real-world conditions is often the most important detail.
The teleoperation caveat in point 2 remains material in 2025 and 2026. Tesla's October 2024 "We, Robot" event drew scrutiny after the company confirmed that the Optimus units mingling with guests were "assisted to some extent" by remote operators in motion-capture suits, although Tesla also says some factory data-collection tasks and later demonstrations run without human intervention. [24] The 1X NEO home robot ships with an explicit human-in-the-loop design: when the robot cannot complete a task autonomously, an owner can schedule an "Expert Mode" session in which a remote 1X operator takes control through virtual reality, and the robot's AI learns from the demonstration. [23]
An honest assessment of the state of humanoid robotics in 2025 reveals both genuine achievements and persistent challenges. [9]
| Category | What works | What still struggles |
|---|---|---|
| Object handling | Tote and bin movement; container unloading; simple pick-and-place of rigid objects | Deformable objects (clothing, food, bags); transparent or reflective objects; very small items |
| Navigation | Structured warehouse aisles; flat factory floors; pre-mapped indoor spaces | Cluttered homes; outdoor terrain with varied surfaces; crowded public spaces |
| Task complexity | Single-step repetitive tasks; scripted multi-step sequences in controlled settings | Open-ended multi-step tasks; novel problem-solving; adapting to unexpected situations |
| Error recovery | Retrying simple grasp failures; recovering from minor balance disturbances | Recovering from novel failures; diagnosing root causes; requesting appropriate help |
| Manipulation precision | Placement within 5 mm tolerance for structured tasks; large object grasping | Fine assembly (screws, connectors); tool use requiring human-level dexterity; force-sensitive tasks |
| Human interaction | Basic voice commands; scripted conversational responses; simple gesture recognition | Natural conversation with context; understanding implicit instructions; reading emotional cues |
| Endurance | 5 to 10 hour shifts for simple, repetitive tasks | Full workday of varied tasks; graceful degradation under component wear; self-maintenance |
The period from 2025 into 2026 saw humanoid robotics move from isolated pilots toward early scaled deployment, with rapid progress in vision-language-action and whole-body foundation models, the first commercial Robots-as-a-Service contracts, and consumer pre-orders.
Several general-purpose robot "brains" advanced the state of the art in 2025. NVIDIA released GR00T N1, described as the first open foundation model for generalist humanoid robots, on March 18, 2025; it is a dual-system vision-language-action model in which a vision-language module (System 2) interprets observations and instructions while a diffusion transformer (System 1) generates motor actions, trained on a mixture of real-robot trajectories, human videos, and synthetic data, and deployed on the Fourier Intelligence GR-1 for bimanual manipulation. [20] NVIDIA subsequently released updated GR00T N1.5 and N1.7 checkpoints. [20] Physical Intelligence published pi-0.5 on April 22, 2025, a VLA that co-trains on heterogeneous data (robot demonstrations from multiple platforms, verbal instructions, subtask labels, object detection, and web data drawn from roughly 100 distinct environments) and demonstrated open-world generalization by cleaning kitchens and bedrooms in homes it had never seen during training. [21] Google DeepMind introduced Gemini Robotics in March 2025 and, in June 2025, Gemini Robotics On-Device, a local VLA that runs on a robot's onboard hardware, is offered for fine-tuning, and can adapt to a new task from as few as 50 to 100 demonstrations. [22] Boston Dynamics and the Toyota Research Institute demonstrated a single Large Behavior Model controlling the electric Atlas across combined manipulation and locomotion on August 20, 2025. [17]
In February 2025 Figure ended its collaboration with OpenAI to develop AI in-house. [18] On February 20, 2025 it unveiled Helix, a "System 1, System 2" VLA that outputs high-rate continuous control of the entire humanoid upper body across a 35-degree-of-freedom action space (wrists, torso, head, and individual fingers); System 2 is a roughly 7-billion-parameter vision-language model running at 7 to 9 Hz, while System 1 is an 80-million-parameter transformer running at 200 Hz, and a single set of weights can run two robots collaborating on a shared task. [18] Figure announced its BotQ manufacturing facility on March 15, 2025, with an initial line capacity of up to 12,000 humanoids per year. [25] Figure unveiled its third-generation robot, Figure 03, in October 2025 as a home-and-commercial platform running Helix, featuring tactile fingertip sensors, palm cameras, and foot-mounted coils for 2 kW wireless inductive charging; it was named to TIME's Best Inventions of 2025. [25] Separately, BMW reported that a Figure 02 pilot at its Spartanburg, South Carolina plant ran for about eleven months, supporting production of more than 30,000 BMW X3 vehicles, moving more than 90,000 components, and covering roughly 1.2 million steps over about 1,250 operating hours of ten-hour shifts before the test concluded; BMW announced a follow-on Physical AI pilot at its Leipzig plant for the summer of 2026 and a new Center of Competence for Physical AI in Production. [19]
Tesla continued scaling Optimus. Elon Musk said production of Optimus version 3 (Gen 3) would begin in 2026, with the new generation moving to a hand with 22 degrees of freedom and tactile fingertip sensors, running on Tesla's AI5 chip, and sharing its computer-vision and neural-network architecture with the company's Full Self-Driving stack. [24] Tesla also stated it would end production of the Model S and Model X to free capacity at its Fremont factory for Optimus, and integrated xAI's Grok model for natural-language interaction. [24] As with the broader industry, Tesla has faced scrutiny over how much of its public demonstrations are teleoperated versus autonomous. [24]
Agility Robotics and the logistics provider GXO signed what both companies called the industry's first multi-year Robots-as-a-Service agreement for humanoid robots, announced in June 2024, deploying the Digit robot together with Agility's Arc cloud platform; by late 2025 Digit had moved more than 100,000 totes at a GXO facility. [27] 1X Technologies opened pre-orders for its NEO home humanoid at roughly 20,000 US dollars to purchase or 499 US dollars per month, with first deliveries to early-access US customers targeted for 2026. [23] Boston Dynamics reported that its all-electric Atlas, shown performing automotive-assembly tasks at CES in January 2026, had its entire 2026 production run committed to parent company Hyundai and to Google DeepMind. [26] In the lower-cost tier, Unitree shipped the H2 full-size humanoid and the compact R1, the latter priced from about 4,900 US dollars and named a TIME Best Invention of 2025. [30]
Investment-bank forecasts diverged widely but pointed to rapid growth. Goldman Sachs raised its 2035 total-addressable-market estimate for humanoid robots roughly sixfold, from about 6 billion to about 38 billion US dollars, citing faster-than-expected progress in supply chains and AI, and projected shipments surpassing 1.4 million units. [28] Morgan Stanley estimated the humanoid market could reach about 5 trillion US dollars by 2050. [28] On the standards side, China's T/CIE 298-2025 intelligence grading framework (above) was followed in early 2026 by a broader national standard system for humanoid robotics and embodied AI, reflecting growing emphasis on safety baselines and shared evaluation criteria as deployments scale. [16]