The Humanoid Compendium

Walking, for the anthropomorphic robotic system, is not a subordinate problem. It is the constitutive one. A platform that cannot reliably walk cannot inhabit the environments for which it is designed; a platform that walks reliably is, in the practical sense of the field, most of the way to being a useful humanoid. The remainder of the problem, manipulation, perception, task planning, is not thereby trivial, but it is significantly downstream of the walking one. The trajectory of humanoid engineering over the sixty years for which the discipline has existed as an ambition is best read as the arc of the walking problem, from static demonstrator to dynamic performer to reliable operational platform.

The present entry treats the walking problem at four levels. The first is the mechanical: what physical arrangement of joints, actuators, and structural elements permits a two-legged system to remain upright and progress forward. The second is the control-theoretic: what mathematical framework describes the coordination of these mechanical elements over the trajectory of a step. The third is the computational: what hardware and software substrate executes the control framework at the rates required. The fourth is the practical: what the current commercial platforms can and cannot do, and where the operational envelope lies as of 2026.

The mechanical substrate

A bipedal humanoid, considered mechanically, is a set of rigid links connected by revolute joints. The links are the body segments (feet, shanks, thighs, pelvis, torso segments where present, arms, hands, head); the joints are the articulations at which one link rotates with respect to another. For the walking problem, the leg joints are the primary subject: the ankle (two or three degrees of freedom depending on the platform), the knee (typically one), and the hip (two or three degrees of freedom). Each joint is actuated, meaning it is driven by a motor or hydraulic element under active control, and each joint is instrumented, meaning its angular position (and typically its angular velocity and applied torque) is measured continuously.

The actuator architecture is a matter of substantial variation across contemporary commercial platforms. Through the mid-2020s, the leading actuator paradigm was the electric quasi-direct-drive motor, an electromagnetic motor with a small reduction ratio, chosen for its combination of torque, backdrivability, and control precision. The paradigm displaced the earlier hydraulic actuator, which had been the dominant choice through the 2010s in platforms including the original Boston Dynamics Atlas, but which had operational disadvantages (leakage, mass, thermal load) that made it unsuitable for the deployment environments of contemporary commercial platforms. Boston Dynamics' 2024 replacement of the hydraulic Atlas with the all-electric variant is the emblematic transition of the shift.

Alongside the electric quasi-direct-drive paradigm, tendon-driven architectures have emerged as an important variant, particularly for the hand. The Tesla Optimus Gen 3, disclosed at the 2026 Abundance Summit, employs a biomimetic tendon-driven hand system with twenty-two degrees of freedom per hand and fifty actuators across the two hands combined. The tendon approach places actuators in the forearm and transmits force to the fingers through cables, an architecture that mirrors the anatomy of the human hand and forearm. Tendon-driven systems are not currently the dominant leg-joint approach, but their advantages at the hand suggest broader application in coming platform generations.

Dynamic balance and the control-theoretic problem

A bipedal humanoid is, at rest, an inverted pendulum. Its centre of mass sits well above its base of support (the feet), and any perturbation displaces it further from equilibrium unless an active control response restores it. The problem of dynamic balance is the problem of maintaining the centre of mass over a base of support that is itself in motion, since walking requires the base to move continuously.

The theoretical treatment of this problem descended, in the modern discipline, from work at the Massachusetts Institute of Technology and Carnegie Mellon University in the 1980s and 1990s. Marc Raibert's Leg Laboratory research, initially at CMU and subsequently at MIT, developed the algorithmic vocabulary of dynamic balance from which contemporary humanoid controllers descend. The vocabulary includes concepts of virtual model control, the zero-moment point, and, more recently, whole-body control frameworks that treat the entire platform as a single coupled system rather than as a set of independently controlled joints.

The practical implementation of these control frameworks requires the resolution of a set of trade-offs that continue to distinguish current commercial platforms. A more agile platform (one capable of dynamic manoeuvres such as running, jumping, or rapid direction changes) requires more aggressive control laws, higher-bandwidth actuators, and a computational substrate capable of executing the control law at the required rate. A more conservative platform (one prioritising steady-state locomotion over agility) can operate with slower actuators and lower-bandwidth control, at correspondingly lower cost.

The computational substrate

A humanoid platform executes its control law on a substrate of onboard computing hardware. Contemporary commercial platforms carry substantial computational resources: the XPeng Next-Gen Iron carries three of XPeng's Turing AI chips, offering effective compute in the two-and-a-half to three thousand TOPS range; Tesla Optimus uses inference silicon derived from the FSD compute platform; other platforms use combinations of GPU, TPU, and custom-designed inference hardware. The compute is used both for the walking control loop (typically at rates from a few hundred to a few thousand Hz) and for the higher-level task-planning and perception loops (typically at rates of tens of Hz).

The compute requirements have declined in the relevant sense over the recent past. The bill of materials for a useful humanoid in 2026 is approximately half what it was in 2024, and the compute component has fallen faster still, as a result of specialised inference silicon becoming available at prices that make it viable for commercial platform integration. This decline is one of the reasons the humanoid category has moved into commercial deployment at all: the platforms are now affordable enough for the addressable market to be commercially interesting.

The operational envelope, 2026

The practical result of the mechanical, control, and computational advances of the recent past is that contemporary commercial humanoid platforms can walk reliably at moderate speeds (typically one to two metres per second) in structured environments, and less reliably at higher speeds and in unstructured environments. The Boston Dynamics electric Atlas is capable of dynamic manoeuvres including precision walking over uneven surfaces and rapid direction changes. The Unitree G1, at a fraction of the price, is capable of moderate-agility locomotion suitable for research applications. Platforms in between the two, including Figure 03, Apptronik Apollo, Sanctuary Phoenix, and 1X NEO, occupy intermediate positions on the agility-cost curve.

The operational envelope narrows significantly when the platform is asked to perform tasks that stress the balance controller: heavy manipulation, carrying loads at range, or operating on soft or slippery surfaces. The envelope also narrows when perception becomes constrained (low light, cluttered environments, adversarial lighting). These are not intrinsic limits of the field but current limits of the commercial state of the art; each is the subject of active research and each is expected to loosen over the coming platform generations.