Kid Cosmo: Entertainment Humanoid Robot

• Kid Cosmo is a child-sized humanoid robot designed for performance-oriented entertainment, uniquely blending character fidelity with dynamic, torque-controlled locomotion.
• It employs innovative mechanical designs such as a unique 5-DoF leg configuration and proprioceptive actuators to achieve high-bandwidth, robust motion.
• The integrated system synchronizes upper-body expressiveness and dynamic gait, addressing live-event constraints, durability, and precise whole-body control.

Kid Cosmo is a child-sized humanoid robot developed as a research platform for entertainment robotics, with the explicit aim of combining robust locomotion, lifelike motion generation, and faithful embodiment of its namesake character from Netflix’s The Electric State. The platform stands 1.45 m tall, weighs 25 kg, and has 28 degrees of freedom. Its design emphasis differs from that of many contemporary humanoids because entertainment prioritizes visuals and form, creating constraints that directly shape mechanical packaging, control architecture, maintenance strategy, and the coupling of expressive upper-body motion with dynamic walking (Liu et al., 16 Aug 2025).

1. Definition and design objective

Kid Cosmo is presented as a humanoid system for performance-oriented entertainment rather than as a purely functional laboratory robot. The paper frames this distinction in terms of a tension between character-based appearance and technical functionality: the robot had to imitate the look and mannerisms of a pre-existing fictional character while remaining capable of torque-control walking and lifelike motion generation. Following worldwide showcases as part of the movie’s press tour, the reported work documents the system architecture, the challenges of a functional entertainment robot, the specific solutions adopted, and initial findings on stability during simultaneous upper and lower body movement (Liu et al., 16 Aug 2025).

This framing is significant because it shifts the optimization target. In conventional humanoid development, morphology is often subordinate to locomotion, manipulation, or sensing objectives. Here, by contrast, external appearance dictated major aspects of internal arrangement. A plausible implication is that Kid Cosmo occupies a boundary region between humanoid robotics, character animation, and live performance systems, where embodiment constraints cannot be treated as a superficial shell added after locomotion design.

2. Mechanical embodiment and morphology

The robot’s shell was made to closely replicate the movie character’s look, and this aesthetic requirement dictated internal mechanical and electronic arrangement. The torso had to be scaled up to 150% from the VFX model to contain required electronics, while hip and knee shell openings were enlarged to accommodate actuators and allow the necessary leg workspace. The resulting platform therefore encodes character fidelity directly into its kinematic and packaging decisions (Liu et al., 16 Aug 2025).

Several structural choices address the mismatch between a visually fragile character and the demands of dynamic locomotion. The large head was constructed from thin 3D-printed nylon and reinforced with an inner carbon fiber frame. Rubber washers between torso and shell add compliance. The boots use spring steel flexures and sliding elements, giving the large feet lifelike, damage-resistant articulation. Magnetic attachments allow the front torso shell to be quickly detached for battery access, which the paper identifies as critical for live entertainment scenarios. The hands use a modified commercial prosthetic, Mand.ro, covered with in-house shells intended to preserve both mobility and appearance.

The joint layout also departs from a standard humanoid template. Kid Cosmo uses unique 5-DoF legs with no ankle roll, explicitly trading off maximal kinematic flexibility for lower distal mass and enhanced dynamic acceleration. The hip joints are configured at 45° and 135° relative to the body axes rather than the typical 0° and 90°, coupling yaw and roll and enabling use of identical, compact actuators. To minimize limb inertia, most leg actuators are placed close to the torso; the knee motor is co-axial with the hip, with torque transferred to the knee through a parallel four-bar linkage. These choices tie morphology to dynamic performance rather than treating appearance and control as separable layers.

3. Actuation and torque-control implementation

A defining technical feature of Kid Cosmo is its reliance on proprioceptive actuators. In the paper, these actuators are described as enabling high-bandwidth torque control essential for rapid, compliant, and robust whole-body movement, and as superior in backdrivability and control smoothness relative to traditional high-transmission actuators (Liu et al., 16 Aug 2025).

The platform uses Westwood Robotics Panda Bear Plus actuators for high-load hip pitch and knee joints, Koala Bear Muscle Build actuators for hip yaw/roll, ankles, and the upper body, and micro gearmotors for wrist and fingers. The knee joint combines a Panda Bear Plus with a 3:1 planetary gearbox to provide higher torque with manageable thermal load.

Component	Specification set
KBMB	285 g, 20:1, 8 Nm stall torque, 1.16 Nm/A
PB Plus	925 g, 10:1, 26.5 Nm stall torque, 1.3 Nm/A
Micro GM	9.5 g, 298:1, 0.32 Nm stall torque, 0.43 Nm/A

All major joints, especially the legs, operate under closed-loop torque control informed by proprioceptive feedback. The final command to each joint uses PD plus feedforward,

$$\mathbf{u} = \mathbf{K}_p(\boldsymbol{\theta}_{\text{des}} - \boldsymbol{\theta}) + \mathbf{K}_d(\boldsymbol{\dot{\theta}}_{\text{des}} - \boldsymbol{\dot{\theta}}) + \boldsymbol{\tau}_{\text{des}},$$

where desired position and velocity come from motion or trajectory planners, $\mathbf{K}_p$ and $\mathbf{K}_d$ are joint PD gains, and $\boldsymbol{\tau}_{\text{des}}$ is the feedforward torque from the model-based controller. Within the paper’s architecture, torque control is not an isolated actuator-level feature; it is the mechanism that permits disturbance rejection, compliant interaction, and dynamic integration of gait with expressive motion.

4. Motion generation, locomotion, and whole-body control

Kid Cosmo’s motion pipeline combines pre-authored expressivity with dynamic locomotion. Motions are pre-recorded by puppeteering the robot’s limbs. Actions and gestures are then sequenced by an operator, using custom or standard controllers, or selected from a motion library. A motion manager uses a finite state machine with REST, ACTION, TRANSITION, STOP, and JOYSTICK states to switch and coordinate behaviors. Smooth transitions are generated with cubic Hermite splines that interpolate joint position and velocity over transitions and ensure $C^1$ continuity (Liu et al., 16 Aug 2025).

Stable walking during simultaneous upper-body motion is handled by a model-based hierarchical control stack. State estimation combines kinematic and contact sensor data with a 3DM-CV7-AHRS IMU through a contact-aided Invariant EKF. For gait generation, the paper uses a Linear Inverted Pendulum model, noting that this model is well matched because Kid Cosmo has a high center of mass and no ankle roll. The equation of motion is

$$\begin{bmatrix} \dot{x}\ \ddot{x} \end{bmatrix} = \begin{bmatrix} 0 & 1\ g/h & 0 \end{bmatrix} \begin{bmatrix} x\ \dot{x} \end{bmatrix},$$

with constant center-of-mass height $h$. Footstep location incorporates angular momentum for robustness:

$$p_{SF}(t_c) = \frac{L_{\text{des}}(T_s) - \cosh(\omega(T_s - t_c))L(t_c)} {mh\omega \sinh(\omega(T_s-t_c))},$$

where $\omega = \sqrt{g/h}$. Swing leg trajectories are represented by a 6th-order polynomial,

$p(\phi) = \sum_{i=0}^{6} c_i \phi^i,$

with $\mathbf{K}_p$0 and endpoint position, velocity, acceleration, plus midpoint height constraints.

Whole-body control is formulated from full-order rigid-body dynamics,

$\mathbf{K}_p$1

where $\mathbf{K}_p$2 contains body pose and joint positions, $\mathbf{K}_p$3 is the mass matrix, $\mathbf{K}_p$4 aggregates Coriolis and gravity terms, and $\mathbf{K}_p$5 denotes contact forces. Tasks are encoded as desired accelerations using PD regulation in task space, and the controller solves a hierarchical implicit WBC formulated as a QP that balances foot, upper-body, and center-of-mass tasks under full-body dynamics, friction cones, and torque limits. This optimization is solved in real time at 1 kHz by the PROXQP solver. Joint position and velocity setpoints for inverse kinematics are obtained with numerical Damped Least Squares. Taken together, these components form the technical basis for synchronized gait and gesture rather than separate motion layers that might interfere destructively.

5. Entertainment-specific constraints and integrated system architecture

The paper identifies four principal challenges for entertainment-oriented whole-body motion: appearance versus function, durability and robustness, fluid simultaneous upper and lower body motion, and wireless interference. These challenges arise from conditions not usually foregrounded in conventional humanoid evaluation: large cosmetic shells, non-humanoid feet, limited sightlines during performance, dense crowds, and live-event communication noise (Liu et al., 16 Aug 2025).

The corresponding solutions are distributed across mechanics, controls, and communication. Mechanical packaging and compliance features in the boots and body prevent impact transmission during dynamic walking and gestures. The hip and knee configuration, along with lightweight design choices, support both faithful appearance and robust gait. Quick-access shell mounting improves maintainability. Wireless reliability is addressed through an in-house 915 MHz wireless controller, introduced because live events with dense crowds and jamming rendered standard controllers inoperable. Control integration is handled by a hierarchical, model-based architecture in which reference trajectory planning for upper and lower body, model-based WBC coordination, and the FSM-based behavior layer are explicitly combined so that gestures and gait remain synchronized rather than destabilizing.

At the systems level, hardware and controllers are described as being designed with equal focus on appearance, performance, and robust, lifelike interaction. Shared-memory-based communication links user interfaces, the motion library and FSM manager, reference planners, and the full-body optimization stack. Custom tools and operator interfaces support both live puppeteering or manual override and scripted sequences. This architecture indicates that Kid Cosmo is neither purely autonomous nor merely teleoperated; instead, it is organized as a hybrid performance system in which authored motion, operator control, and model-based stabilization coexist.

6. Experimental findings and place within entertainment robotics

The reported experiments focus on compliance, torque behavior, disturbance rejection, and simultaneous walking with gestures. Compliance testing showed boot-shell stiffness as low as 0.098–0.147 N/mm; even at 16 mm displacement, force never exceeded 2.4 N, which is less than 1% of robot weight. Body shells were stiffer, at approximately 5.8 N/mm, but were reported as sufficient to absorb rare impacts. Torque measurements for the 45°/135° hip configuration yielded well-balanced output, with hip yaw at 3.93 Nm average and hip roll at 3.98 Nm average; the reported torque data showed smooth transitions and minimal mechanical backlash (Liu et al., 16 Aug 2025).

For disturbance rejection, lateral and sagittal pushes displaced the body center of mass by as much as 8 cm, after which the robot returned to a stable position in less than 2 s, with settling velocity below 0.01 m/s and stepping as needed. In simultaneous locomotion and expressivity tests, Kid Cosmo walked at 0.16 m/s while executing right-arm gestures including idle, thumbs-up, and waving, without losing balance; the average velocity tracking error was 0.012 m/s. The FSM and motion blending produced smooth gesture transitions both while stationary and during locomotion.

These results are used in the paper to support a broader claim about entertainment robotics: Kid Cosmo demonstrates the viability of performance-oriented humanoid robots that prioritize both character embodiment and technical functionality. A common misconception is to equate entertainment humanoids with teleoperated or animatronic systems, or with “slow, careful” gait. Kid Cosmo is positioned explicitly against that baseline by combining faithful, lifelike embodiment of a fictional character with robust, dynamic, torque-controlled locomotion and real-time upper- and lower-body expressiveness. This suggests an emerging design category in which artistic fidelity, maintainability, and disturbance-tolerant whole-body control are treated as co-equal system requirements rather than sequential design stages.