The Open Humanoid: Why We Are Building a Robot From First Principles

---

The Problem with Closed Systems

In February 2026, Boston Dynamics announced that its electric Atlas humanoid had entered production and begun autonomous operation in commercial facilities. The robot stands approximately 1.5 meters tall, weighs 89 kilograms, features 28 degrees of freedom, and can perform dynamic movements that were science fiction a decade ago. Tesla claims its Optimus robot will achieve commercial deployment by 2027 at a target price of $20,000-$30,000. Figure AI has completed a ten-month pilot at BMW’s Spartanburg plant, where Figure 02 worked daily ten-hour shifts on the X3 body shop, moving over 90,000 sheet metal components across 1,250 operating hours. These are remarkable achievements. They are also scientifically useless. Not because the engineering lacks merit, but because the systems are closed. The actuator specifications, control algorithms, sensor fusion architectures, and failure modes remain proprietary. When Atlas executes a recovery maneuver after a push perturbation, the research community cannot learn how. When Optimus manipulates objects with its 40-degree-of-freedom hands, no one outside Tesla can reproduce the torque curves or grasp planning logic. When Figure 02 achieves reliable pick-and-place in a real factory environment, the safety protocols and fault tolerance mechanisms stay behind corporate walls. Science advances through reproducibility. A result that cannot be independently verified is not science but demonstration. The humanoid robotics field has accumulated impressive demonstrations. What it lacks is a reproducible body of engineering knowledge that allows researchers, educators, and entrepreneurs to build upon established foundations rather than reinventing them in isolation. This series proposes a different approach: building an autonomous humanoid robot entirely in the open, from specification to simulation.

graph LR
    subgraph Closed["Closed Systems (Boston Dynamics, Tesla)"]
        direction TB
        P1[Proprietary Hardware] --> P2[Locked SDK]
        P2 --> P3[No community contribution]
        P3 --> P4[High cost barrier]
    end
    subgraph Open["Open Humanoid Approach"]
        direction TB
        O1[Open Specification] --> O2[Open Hardware BOM]
        O2 --> O3[Open Software Stack]
        O3 --> O4[Community-driven iteration]
    end
    style Closed fill:#fee2e2,stroke:#dc2626
    style Open fill:#d1fae5,stroke:#10b981

What We Are Building

Over twenty articles, we will design a complete humanoid robot system. Not a conceptual robot or a theoretical framework, but a fully specified engineering artifact with defined constraints, performance targets, and validated simulation models. The final deliverable is a simulation room with two autonomous robots that walk, balance, perceive their environment, communicate with each other, and recover from perturbations. The specification-driven methodology means every subsystem carries explicit constraints: mass budgets, power budgets, volume allocations, cost targets, and performance metrics. When we specify a locomotion subsystem, we define the target gait speed (1.5 m/s minimum), degrees of freedom (12 for lower body), balance recovery time (under 500ms), and ground clearance during swing phase. When we specify a manipulation subsystem, we define grip force ranges, finger degrees of freedom, payload capacity, and positioning accuracy. This is not a textbook approach where equations precede implementation. It is not a hobbyist approach where components accumulate without integration testing. It is the engineering approach: requirements first, then architecture, then detailed design, then validation. The difference is that everything happens in public view.

The Current Landscape: A Technical Survey

Before specifying what we will build, we must understand what already exists. The humanoid robotics field in 2026 features several mature platforms with documented capabilities.

Boston Dynamics Atlas

The Atlas platform underwent a fundamental redesign in 2024, transitioning from hydraulic to fully electric actuation. The production version entering commercial deployment in 2026 stands approximately 1.5 meters tall and weighs 89 kilograms. Secondary sources report approximately 28 degrees of freedom across the kinematic chain. Atlas represents the state of the art in dynamic locomotion. The robot executes running gaits, jumping maneuvers, and recovery motions that demonstrate sophisticated whole-body control. However, the system remains research-oriented in its deployment philosophy. Boston Dynamics has integrated NVIDIA’s Jetson Thor platform for next-generation units, suggesting substantial onboard compute for real-time control and perception. The primary limitation for scientific purposes is complete opacity. No technical publications describe the actuation architecture, control hierarchy, or state estimation approach. The system exists as capability demonstration rather than reproducible engineering artifact.

Tesla Optimus

Tesla’s humanoid robot stands 1.73 meters (5 feet 8 inches) tall and weighs approximately 57 kilograms, making it nearly 30% lighter than Atlas. The Gen 2 design features 40 degrees of freedom with particular emphasis on hand articulation for manipulation tasks. Tesla deploys Optimus internally at its own factories, with commercial availability projected for 2026-2027 at a $20,000-$30,000 price point. Optimus exemplifies the automotive manufacturing approach to robotics: vertical integration, aggressive cost reduction, and scaled production targeting. The actuator designs reportedly leverage Tesla’s expertise in electric motor manufacturing from its vehicle powertrain business. However, independent analysis in early 2026 raised questions about autonomous capability. Reports suggest Optimus demonstrations may rely more heavily on teleoperation than autonomous control, creating uncertainty about the system’s actual decision-making architecture.

Unitree H1 and G1

Chinese manufacturer Unitree has emerged as a significant force in accessible humanoid robotics. The H1 model stands 180 centimeters tall, weighs approximately 47 kilograms, and holds the world record for humanoid running speed at 3.3 meters per second. The robot features peak torque density of 189 N.m/kg in its joint actuators and has shipped over 5,500 units to research institutions and commercial customers. The smaller G1 model stands 127 centimeters (4.2 feet), weighs 35 kilograms, and offers 23-43 degrees of freedom depending on configuration. Priced starting around $16,000, the G1 represents the current price-performance frontier for research-grade humanoid platforms. The robot runs on a 9,000 mAh removable battery and incorporates LiDAR and depth cameras for perception. Unitree’s approach favors accessibility over capability completeness. The robots use electric actuators throughout and support standard robotics development frameworks including ROS. While more open than competitors, Unitree still restricts access to low-level firmware and actuator control implementations. Notably, Unitree demonstrated extreme-environment capability in early 2026 when the G1 completed 130,000 steps in -53 degrees Fahrenheit conditions, validating locomotion algorithms under thermal stress.

Figure AI

Figure AI has pursued aggressive commercial deployment with automotive manufacturers. The Figure 02 robot completed a ten-month pilot at BMW’s Spartanburg facility, working daily ten-hour shifts in the X3 body shop. Over the pilot period, the robot handled over 90,000 sheet metal components and logged approximately 1,250 operating hours. BMW announced expanded deployment in early 2026, bringing humanoid robots to its Leipzig plant in Germany as the first European deployment. Figure is transitioning to its third-generation platform, Figure 03, which features completely redesigned wrist electronics that eliminate distribution boards and dynamic cabling in the forearm subsystem. The Figure program demonstrates that humanoid robots can achieve commercial viability in structured industrial environments. However, the systems operate in closely controlled conditions with predefined tasks rather than general-purpose autonomy.

Agility Robotics Digit

Agility Robotics takes a different approach with Digit, a robot standing 5 feet 9 inches tall and weighing 143 pounds (65 kilograms). Digit’s distinctive bird-leg design improves mobility and energy efficiency while supporting a 16-kilogram (35-pound) payload capacity. The company operates the world’s first factory dedicated to humanoid robot manufacturing in Salem, Oregon. Digit focuses on warehouse and logistics applications: tote loading and unloading, line feeding, and material transport. Toyota Canada announced in early 2026 that seven Digit robots would deploy at its Woodstock RAV4 plant. Digit demonstrates commercial viability in the logistics domain but trades manipulation capability for locomotion efficiency. The robot’s arms have limited degrees of freedom compared to other platforms.

Why Open Specification Matters

The platforms surveyed above share a common characteristic: they cannot be reproduced. A researcher reading their specifications cannot build a comparable system. An educator cannot explain the engineering tradeoffs to students. An entrepreneur cannot adapt the designs for new applications. This matters for three reasons. Scientific Progress Requires Reproducibility. Koseki et al. (2026) publish in the Journal of the Royal Society Interface their work on human-inspired bipedal locomotion, connecting neuromechanics to mathematical modeling and robotic applications. Their contribution advances science because other researchers can implement their models, test their predictions, and build upon their foundations. When Boston Dynamics demonstrates a backflip, the spectacle impresses but teaches nothing that others can verify. Education Requires Transparency. Engineering education depends on worked examples with visible tradeoffs. Students learn motor selection by seeing how torque requirements, speed ranges, thermal limits, and cost constraints interact in actual design decisions. Closed systems provide no educational value beyond existence proofs. Innovation Requires Foundations. The open source software movement demonstrated that shared infrastructure accelerates innovation by allowing effort concentration on novel contributions rather than reinventing standard components. Humanoid robotics needs equivalent shared foundations: reference architectures, validated simulation models, and documented failure modes. Recent academic work supports the open methodology. Research published in 2025-2026 on deep reinforcement learning for bipedal locomotion demonstrates that simulation-to-reality transfer can achieve robust walking using only proprioceptive feedback when trained with appropriate curriculum randomization. Hierarchical control frameworks for humanoids with supernumerary limbs show how to decompose complex whole-body control problems into tractable subproblems. Work on closed kinematic chains in bipedal designs addresses the mechanical complexity of parallel mechanisms common in practical robots. These contributions exist because researchers published methods, not just results. Our series applies the same principle to integrated system design.

The Spec-Driven Approach

Every article in this series references a master specification schema that defines how subsystem requirements are documented:

subsystem: [name]
version: 0.1
dependencies: [list of other subsystems]
constraints:
  mass_budget_kg: [value]
  power_budget_w: [value]
  volume_mm: [L x W x H]
  cost_usd: [target BOM cost]
performance_targets:
  [metric]: [value with unit]
open_challenges:
  - [challenge 1]
  - [challenge 2]

The schema enforces discipline. Every specification must state its mass budget (the robot cannot exceed 80 kilograms total), power budget (the robot must operate for 60+ minutes), volume constraints (components must fit the humanoid form factor), and cost targets (we aim for research-accessible pricing). Global constraints apply to all subsystems:

• Total mass: 80 kilograms maximum
• Height: 160-180 centimeters
• Battery life: greater than 60 minutes under normal operation
• Operating temperature: 0-40 degrees Celsius
• IP rating: IP54 minimum (dust protected, splash resistant)
• Emergency stop: hardware and software, under 100 milliseconds response
• Communication: onboard WiFi 6 and Bluetooth 5.2, optional 5G

These constraints derive from practical requirements. An 80-kilogram robot can be handled by two technicians, fits through standard doorways, and stays within reasonable actuator sizing. The 60-minute battery target enables useful work cycles. IP54 protection allows indoor operation with incidental liquid exposure. Sub-100ms emergency stop meets industrial safety standards.

Series Roadmap

The twenty articles in this series cover the complete system:

graph TD
    A1[Article 1: Introduction] --> A2[Article 2: System Specification]
    A2 --> A3[Article 3: Locomotion]
    A2 --> A4[Article 4: Actuation]
    A2 --> A5[Article 5: Structure]
    A2 --> A6[Article 6: Power]
    A2 --> A7[Article 7: Compute]
    A3 --> A8[Article 8: Vision]
    A4 --> A8
    A8 --> A9[Article 9: Sensor Fusion]
    A5 --> A10[Article 10: Manipulation]
    A4 --> A10
    A7 --> A11[Article 11: Speech]
    A9 --> A12[Article 12: Navigation]
    A10 --> A13[Article 13: Force Control]
    A6 --> A14[Article 14: Safety]
    A7 --> A15[Article 15: Control Architecture]
    A15 --> A16[Article 16: Multi-Robot Communication]
    A12 --> A17[Article 17: Simulation Environment]
    A16 --> A17
    A17 --> A18[Article 18: Integration]
    A14 --> A18
    A18 --> A19[Article 19: Assembly Guide]
    A18 --> A20[Article 20: Simulation Room]

Article 1 (this article): Introduction and landscape survey. Article 2: Complete high-level specification with all constraint budgets and subsystem interfaces. Article 3: Bipedal gait and balance control, including zero-moment point calculation, center of mass tracking, and push recovery. Article 4: Motor selection, torque requirements, and degree-of-freedom allocation across the kinematic chain. Article 5: Structural design including material selection (carbon fiber, aluminum, TPU), stress analysis, and mass optimization. Article 6: Battery chemistry, capacity, battery management system design, and power distribution. Article 7: Onboard compute requirements, real-time operating system selection, and latency budgets. Article 8: Computer vision pipeline including depth sensing, object detection, and simultaneous localization and mapping. Article 9: Sensor fusion architecture combining IMU, force sensors, joint encoders, and vision. Article 10: Hand and manipulation design including finger degrees of freedom, grip force, and grasping strategies. Article 11: Speech interface including automatic speech recognition, text-to-speech, and onboard language model integration. Article 12: Navigation including path planning, obstacle avoidance, and terrain adaptation. Article 13: Force control and haptics for compliant manipulation and contact-rich tasks. Article 14: Safety systems including fall detection, emergency stop implementation, and IP54 sealing. Article 15: Real-time control architecture including loop rates and latency budgets across the control hierarchy. Article 16: Multi-robot communication protocols and distributed decision-making. Article 17: Simulation environment using physics engines for digital twin validation. Article 18: Integration and testing methodology for connecting all subsystems. Article 19: Bill of materials and sourcing guide for physical implementation. Article 20: Final simulation room demonstration with two autonomous robots.

graph TD
    subgraph Perception
        V[Vision & Depth Camera] --> SF[Sensor Fusion]
        IMU[IMU] --> SF
        FT[Force-Torque Sensors] --> SF
    end
    subgraph Control
        SF --> Nav[Navigation]
        SF --> LC[Locomotion Controller]
        Nav --> MC[Motion Coordinator]
        LC --> MC
    end
    subgraph Actuation
        MC --> Legs[Leg Motors x12 DOF]
        MC --> Arms[Arm Motors x14 DOF]
        MC --> Hands[Hand Actuators x12 DOF]
    end
    subgraph Intelligence
        LLM[Onboard LLM] --> Nav
        Safety[Safety Monitor] --> MC
    end
    style Perception fill:#dbeafe
    style Control fill:#fef9c3
    style Actuation fill:#d1fae5
    style Intelligence fill:#fce7f3

The Simulation Room Vision

The series culminates in a Three.js browser simulation showing two humanoid robots operating in a shared space. Each robot:

• Walks with dynamically stable bipedal gait
• Maintains balance through active center-of-mass control
• Perceives the environment through simulated vision
• Communicates state information with the other robot
• Recovers from simulated perturbations

The simulation runs in real-time in a web browser, demonstrating that the specified systems achieve the stated performance targets within their computational budgets. The simulation is not the goal; validation against specification is the goal. The simulation makes that validation visible and reproducible. The simulation room represents the final proof that open, spec-driven engineering can produce capable robotic systems. When both robots walk, communicate, and adapt without hidden proprietary components, the methodology is validated.

Conclusion

The humanoid robotics field has demonstrated remarkable capabilities but failed to establish reproducible engineering foundations. The Open Humanoid series addresses this gap through systematic, public specification of a complete autonomous robot system. We are not building a better robot than Atlas or Optimus. We are building an understandable robot. The value lies not in capability supremacy but in transparency that enables scientific progress, education, and innovation. The next article presents the complete high-level specification: every constraint, every interface, every performance target. The engineering begins there.

References

1. Koseki, S., Hayashibe, M., & Owaki, D. (2026). Human-inspired bipedal locomotion: from neuromechanics to mathematical modelling and robotic applications. Journal of the Royal Society Interface, 23(235), 20250662.

1. Boston Dynamics. (2026). Atlas enters production for commercial deployment. The Register, January 6, 2026.

1. BMW Group & Figure AI. (2025). Humanoid robots complete trial project at BMW assembly plant. Assembly Magazine, November 25, 2025.

1. Unitree Robotics. (2026). Unitree G1 achieves 130,000 steps in extreme cold conditions. eWeek, February 2026.

1. Toyota Canada. (2026). Toyota Canada confirms 2026 rollout of Agility’s Digit robots in Woodstock factory. Yahoo Finance, February 2026.

1. arXiv. (2026). Robust humanoid walking on compliant and uneven terrain with deep reinforcement learning. arXiv:2504.13619.

1. arXiv. (2025). A hierarchical framework for humanoid locomotion with supernumerary limbs. arXiv:2512.00077.

1. arXiv. (2025). Robust RL control for bipedal locomotion with closed kinematic chains. arXiv:2507.10164.

1. Radosavovic, I., et al. (2026). Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469v3. Retrieved March 2026 from https://arxiv.org/abs/2402.19469.

1. Fu, Z., et al. (2026). HumanPlus: Humanoid Shadowing and Imitation from Humans. arXiv preprint arXiv:2406.10454. Retrieved March 2026 from https://arxiv.org/abs/2406.10454.