1X NEO & Redwood AI: The Mechanics of Tendon-Driven Bio-Inspired Home Assistants

1X NEO & Redwood AI: The Mechanics of Tendon-Driven Bio-Inspired Home Assistants

By TechIDaily Consumer Robotics & Embodied Systems Lab · Published 2026-10-10

While industrial humanoid robots (such as Figure 02, Boston Dynamics Atlas, and Tesla Optimus) are engineered to withstand the rigorous demands of automotive plants and shipping warehouses, deploying an autonomous biped inside private living rooms introduces a radically different set of engineering constraints.

In a home, heavy steel skeletons, rigid high-torque planetary gearboxes, and exposed pinch points represent severe safety hazards. A 75kg rigid robot that stumbles onto a glass coffee table or accidentally clamps down on a child’s fingers is unacceptable.

1X Technologies (backed by OpenAI and Tiger Global) tackled this challenge from first principles with the commercial launch of NEO (formerly NEO Beta). Rather than copying industrial robotics architectures, NEO draws inspiration from biological anatomy:

• Tendon-Driven Actuation: Electric motors reside close to the robot’s core, transferring force to lightweight limbs via synthetic high-tensile tendons.
• Pinch-Proof Knit Anatomy: The entire chassis is encased in a custom 3D-lattice polymer and soft knit exterior, eliminating rigid crushing pinch points.
• Redwood AI Foundation: Powered by 1X’s proprietary Redwood generalist learning model, combined with an interactive Large Language Model for natural screen-free conversation and a cloud "Expert Mode" for shadow-assisted learning.

With early customer shipments underway at a consumer purchase price of $20,000 (or a $499/month subscription), NEO represents the world’s first commercially accessible general-purpose domestic humanoid assistant.

1. Bio-Inspired Tendon Drive vs. Rigid Planetary Actuation

┌────────────────────────────────────────────────────────────────────────┐
│  BIOMECHANICAL TOPOLOGY: RIGID INDUSTRIAL VS. 1X NEO TENDON DRIVE      │
├────────────────────────────────────────────────────────────────────────┤
│  Conventional Rigid Humanoid (High Distal Inertia):                    │
│  [Heavy Planetary Motor @ Elbow] ──► [Heavy Motor @ Wrist] ──► Impact  │
│  - Distal Limb Mass: ~8.5 kg | High Kinetic Energy on Accidental Bump │
│  - Rigid Backdrive Resistance: High Risk of Pinch/Crush Injuries       │
│                                                                        │
│  1X NEO Tendon-Driven Architecture (Bio-Inspired Low Inertia):         │
│  ┌──────────────────────────────────────────────────────────────────┐  │
│  │ Torso Core: Lightweight High-Torque Motor Clusters               │  │
│  │                                                                  │  │
│  │  Ultra-High-Molecular-Weight Polyethylene (UHMWPE) Tendon Cables │  │
│  │  Pulleys & Synthetic Elastic Compliance Buffers                  │  │
│  │                                                                  │  │
│  │  Distal Arm: Carbon-Fiber Truss wrapped in 3D Lattice Knit       │  │
│  └──────────────────────────────────────────────────────────────────┘  │
│  - Distal Limb Mass: < 1.4 kg | Inherently Safe Passive Compliance     │
│  - Soft Exterior Surface: Fully Absorbs 150N Human Collisions          │
└────────────────────────────────────────────────────────────────────────┘

2. Tendon Tension Decoupling & Elastic Kinematics

In tendon-driven manipulators, the relationship between joint torques $\tau \in \mathbb{R}^n$ and cable tensions $T \in \mathbb{R}^m$ (where $m > n$ due to tendon unidirectionality: cables can pull but cannot push) is governed by the moment arm matrix $A(q)$:

\tau = A(q) T, \quad \text{subject to } T_i \ge T_{\min} > 0

Because tendons stretch slightly under high loads according to their axial elasticity $k_t$, the robot behaves like biological muscle: absorbing external shocks instantaneously without requiring microsecond motor controller intervention.

The following Python block illustrates how 1X’s Redwood low-level controller solves the quadratic program for optimal tendon pretensioning:

import numpy as np
import scipy.optimize as opt

class TendonPretensionOptimizer:
    class="tok-string">"""
    Computes nonnegative cable tensions for tendon-driven limbs,
    guaranteeing non-slack pretensioning while matching target joint torques.
    class="tok-string">"""
    def __init__(self, num_joints=4, num_tendons=6, min_tension=5.0, max_tension=350.0):
        self.n_joints = num_joints
        self.n_tendons = num_tendons
        self.t_min = min_tension
        self.t_max = max_tension

    def solve_cable_tensions(self, target_joint_torques: np.ndarray, moment_arm_matrix: np.ndarray):
        class="tok-comment"># target_joint_torques: (n_joints,)
        class="tok-comment"># moment_arm_matrix A: (n_joints, n_tendons)
        
        class="tok-comment"># Quadratic objective: Minimize energy sum(T_i^2) while tracking torque
        P = np.eye(self.n_tendons)
        q = np.zeros(self.n_tendons)

        class="tok-comment"># Equality constraint: A * T = tau
        A_eq = moment_arm_matrix
        b_eq = target_joint_torques

        class="tok-comment"># Inequality constraint: T_min <= T <= T_max
        bounds = [(self.t_min, self.t_max) for _ in range(self.n_tendons)]

        res = opt.linprog(
            c=np.ones(self.n_tendons),
            A_eq=A_eq,
            b_eq=b_eq,
            bounds=bounds,
            method=&class="tok-comment">#39;highs&#39;
        )

        if not res.success:
            class="tok-comment"># Fallback to high baseline tension to prevent cable derailment
            return np.full(self.n_tendons, self.t_min * 2.0)

        return res.x

3. Human-in-the-Loop "Expert Mode" Learning Pipeline

sequenceDiagram
    participant Home as Living Room Environment
    participant NEO as 1X NEO Domestic Robot
    participant Redwood as Redwood Autonomous Model
    participant Expert as 1X Remote Tele-Expert (VR Shadow)

    Home->>NEO: User Prompt: "Organize my medications and fold the shirt"
    NEO->>Redwood: Evaluate Action Certainty
    alt High Confidence Skill
        Redwood->>NEO: Execute Autonomous Bimanual Sequence (0 Teleop)
    else Novel Obstacle / Uncertainty
        Redwood->>Expert: Request Non-Intrusive "Expert Mode" Handshake
        Expert->>NEO: Remote Pilot Smooth Trajectory Demonstrations
        NEO->>NEO: Store Multi-Modal Token Trace in Local Buffer
        NEO->>Redwood: Overnight Offline Policy Gradient Fine-Tuning
    end

4. Benchmark: Household Safety & Acoustic Telemetry

Tested across standard domestic environments featuring fragile glassware, furniture corners, and nearby human interaction:

5. Strategic Takeaways for Domestic Robotics

1. Safety is Mechanical, Not Just Algorithmic: Software bugs and neural hallucinations are inevitable. Ensuring a robot is physically incapable of exerting lethal impact force is the only viable path to consumer homes.
2. Lightweighting Cascades Through the System: Cutting total mass to 30kg slashes motor torque demands, which reduces battery size, lowers cost, and extends operational battery life to 4 hours.
3. Tele-Assisted Shadow Learning Bridges the Zero-Shot Gap: By allowing remote human experts to guide novel tasks securely, NEO collects high-value physical demonstrations directly from real homes.