asRoBallet: RL-Controlled Humanoid Ballbot

• asRoBallet is a humanoid ballbot system that closes the Sim2Real gap using friction-aware reinforcement learning and high-fidelity MuJoCo simulation.
• The platform employs a PPO-based control framework to achieve precise velocity tracking and station keeping even under dynamic friction variations.
• Its subtractive-reconfigured 49-DoF hardware combined with an iOS control ecosystem democratizes expressive robotic performance and underactuated mobile manipulation.

Searching arXiv for the asRoBallet paper and closely related robotic choreography/control work. asRoBallet is a humanoid ballbot system introduced as “the first successful deployment of reinforcement learning (RL) on a humanoid ballbot hardware,” with the stated objective of closing the Sim2Real gap for underactuated and nonholonomic control in ballbots through friction-aware modeling, domain-randomized RL, and a low-cost hardware platform derived from an overconstrained quadruped (Wan et al., 27 Apr 2026). The system integrates a high-fidelity MuJoCo model of wheel–sphere–ground contact, a PPO-based control framework for velocity tracking and station keeping, a subtractive-reconfigured hardware embodiment with 49 DoF, and a generalized iOS interface for single-operator orchestration of expressive humanoid maneuvers (Wan et al., 27 Apr 2026). In the broader literature on robotic choreography and expressive robotic systems, asRoBallet sits at the intersection of underactuated mobile robotics, sim-to-real transfer, and movement-centered human-robot interaction (LaViers et al., 2017), while differing from prior robot dance systems that focused on industrial arms (Cuan et al., 2024), humanoid dance generation (Augello et al., 2017), or dance-training interaction (Granados et al., 2018).

1. System identity and research problem

asRoBallet addresses a long-standing control problem associated with ballbots, which the source describes as “a canonical benchmark for underactuated and nonholonomic control,” and identifies as being limited by “a reality gap in complex friction models for wheel-sphere-ground interactions” (Wan et al., 27 Apr 2026). The paper states that current literature had already demonstrated “successful handling of 3D balancing with LQR and MPC,” but that “transitioning to actual hardware for a humanoid ballbot using RL is currently hindered by critical gaps in contact modeling, actuator latency & jitter, and safe hardware exploration” (Wan et al., 27 Apr 2026).

The system is presented as a tightly coupled combination of three components: “(1) a very high-fidelity MuJoCo simulation of underactuated spherical dynamics with explicit discrete roller mechanics, (2) a friction-aware reinforcement-learning framework that masters coupled rolling, lateral and torsional friction channels, and (3) a hardware embodiment built by subtractive reconfiguration of a quasi-direct-drive quadruped and controlled via a consumer-grade iOS ecosystem” (Wan et al., 27 Apr 2026). This framing is central to the asRoBallet contribution: the work is not only an RL controller for balancing, but an end-to-end platform spanning simulator fidelity, learning design, hardware embodiment, and teleoperation interface.

A plausible implication is that the name “asRoBallet” serves a double role. It identifies the specific humanoid ballbot platform of (Wan et al., 27 Apr 2026), while also resonating with a wider body of robot performance and choreographic research in which movement expressivity, audience interpretation, and robot embodiment are treated as technical design variables rather than merely artistic superstructure (LaViers et al., 2017, Cuan et al., 2024).

2. High-fidelity MuJoCo model and frictional contact structure

The simulation model is designed to reproduce contact phenomena that simpler ballbot abstractions omit. In MuJoCo, asRoBallet is modeled “as a floating-base multibody comprising the torso, three ETH-type omni-wheel drive units, and a hollow steel sphere coated with SBR rubber” (Wan et al., 27 Apr 2026). Each omni-wheel is represented as “a rigid ring fitted with 12 free-spinning rollers, alternating between load-bearing and spacer rollers, connected by hinge joints” (Wan et al., 27 Apr 2026). According to the source, this “discrete geometric fidelity is essential to reproduce lateral compliance, parasitic vibrations during roller handovers, and contact discontinuities” (Wan et al., 27 Apr 2026).

At the sphere-ground interface, the model enforces “full six-dimensional contact: a normal force $N$, two tangential sliding directions, torsional spin, and two rolling resistance channels” (Wan et al., 27 Apr 2026). The rolling resistance is specified as

$F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$

with the additional damping term

$$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$

which is used “to capture parasitic vibrations during roller handover” (Wan et al., 27 Apr 2026). Tangential directions use Coulomb friction,

$F_{t} = \mu_{\mathrm{slide}\,N\,\mathrm{sign}(v_{t})\,,$

and torsional friction is modeled as

$M_{\mathrm{torsion} = \begin{cases} \mu_{\mathrm{torsion}\,N &\text{if }\omega_{z}\neq0,\ \le \mu_{\mathrm{torsion}\,N &\text{if }\omega_{z}=0 \end{cases}$

to represent stick-slip discontinuity (Wan et al., 27 Apr 2026).

The ball–roller interface uses “standard three-dimensional contact with a soft penetration law”

$\mathrm{solimp}=(0.85,0.99,0.003)$

so that “the rubber layer stiffens over 3 mm, reproducing hysteresis and anisotropic slip arising purely from roller compliance” (Wan et al., 27 Apr 2026). Motor-side actuator friction is identified experimentally through

$$I_\mathrm{m}\,\ddot\theta = \tau - D_v\,\dot\theta - D_c\,\mathrm{sign}(\dot\theta),$$

where “$D_c\approx130$ mNm is the Coulomb breakaway torque measured experimentally” and the resulting dead-zone torque “creates micro-oscillatory limit cycles in station keeping” (Wan et al., 27 Apr 2026).

The simulator configuration is also specified in detail: MuJoCo uses the “elliptic friction cone (cone=‘elliptic’), a stiff constraint impedance ratio (impratio=10), and the Newton solver with noslip_iterations=1 to faithfully resolve stick-slip transitions at 0.2 ms internal steps (frame_skip=5 for a 0.01 s control period)” (Wan et al., 27 Apr 2026). The paper’s central claim is that explicit tribological modeling of rolling, sliding, torsional friction, and roller handover dynamics is necessary for zero-shot transfer on this class of underactuated platform (Wan et al., 27 Apr 2026).

3. Friction-aware reinforcement learning framework

The RL framework trains “two base policies—velocity tracking and station keeping—over this high-fidelity model using a standard on-policy actor-critic algorithm (PPO)” (Wan et al., 27 Apr 2026). The action space is

$a_t\in\mathbb R^3$

and consists of “direct torque commands $\tau$ to the three omni-wheels” (Wan et al., 27 Apr 2026).

For velocity tracking, the observation is

$F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$0

and includes “body-frame linear velocity $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$1, commanded velocity $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$2, tilt quaternion $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$3, gyro rates $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$4, and previous action” (Wan et al., 27 Apr 2026). For station keeping, the observation appends “world-frame position error $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$5 and heading error $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$6” (Wan et al., 27 Apr 2026).

The reward combines exponential tracking terms,

$F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$7

penalties

$F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$8

and an “uprightness constraint $F_{\mathrm{roll} = \mu_{\mathrm{roll}\,N$9” (Wan et al., 27 Apr 2026). For station keeping, the velocity-tracking terms are replaced with “position and heading terms $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$0” (Wan et al., 27 Apr 2026). The total return is

$$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$1

and is optimized “with standard PPO clipping and an $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$2 weight penalty” (Wan et al., 27 Apr 2026).

A distinctive feature of the method is friction-aware domain randomization. The paper states that “friction parameters $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$3 and IMU noise $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$4 are domain-randomized each episode” according to

$$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$5

(Wan et al., 27 Apr 2026) The source further notes that “a virtual low-damped ball joint between body and sphere enforces contact throughout training; ablations show that without this coupling the reward landscape is too sparse and training fails to converge” (Wan et al., 27 Apr 2026).

This design differentiates asRoBallet from earlier robot dance generation systems that focused on latent-variable movement synthesis from music (Augello et al., 2017) or choreographic motion authoring (LaViers et al., 2017). Here, the central learning problem is robust control under highly coupled contact uncertainty rather than sequence generation for stylistic novelty.

4. Hardware embodiment and iOS control ecosystem

The hardware platform is produced by “subtractive reconfiguration” of the open-source “asOverDog quadruped” (Wan et al., 27 Apr 2026). The “overconstrained four-leg drivetrains become three 5.5 Nm RobStride05 actuators driving ETH omni-wheels around the sphere,” while “the higher-torque 17 Nm RobStride01 actuators reassemble into a 2-DoF head and two 4-DoF arms” (Wan et al., 27 Apr 2026). The robot retains “the original NUC + Jetson Orin computational core, a 2 kHz IMU and current-sensing quasi-direct-drive actuators for intrinsic proprioception” (Wan et al., 27 Apr 2026).

The structural design is organized to reduce inertia about the center of mass: “A T-shaped sheet-metal shoulder frame and 3D-printed connectors minimize inertia about the CoM” (Wan et al., 27 Apr 2026). The paper reports that the “final robot has 49 DoF (13 active, 36 passive) and costs an order of magnitude less than custom ballbot drivetrains” (Wan et al., 27 Apr 2026). The passive DoF reflect the compliant and rolling contact elements of the system rather than purely articulated anthropomorphic joints.

The whole-body interface uses commodity mobile hardware. An “iPhone 12 Pro mounted on the spine runs ARKit VIO at 100 Hz, delivering drift-free 6 DoF pose with sub-5 % relative position error and ~1.2 ms peer-to-peer latency via Wi-Fi Aware” (Wan et al., 27 Apr 2026). A “custom iOS app unifies three control modalities—from touchscreen joystick to wearable IMU suit to markerless whole-body pose tracking across iPhone, Watch and AirPods—allowing a single operator to teleoperate asRoBallet’s base and upper body in intuitive, expressive manners” (Wan et al., 27 Apr 2026).

This hardware-software integration places asRoBallet within a broader trend in robot performance systems that couple control architecture to choreographic usability. Comparable concerns appear in industrial-arm performance systems using teach mode and force-mode improvisation (Cuan et al., 2024), and in movement-centered interface design for expressive robotic systems (LaViers et al., 2017). In asRoBallet, however, the control interface is directly linked to an underactuated mobile base whose primary stabilization policy is learned rather than scripted (Wan et al., 27 Apr 2026).

5. Experimental results and zero-shot Sim2Real transfer

The simulation results are reported under “four increasing levels of difficulty (random friction, random tilt, random arm payload)” (Wan et al., 27 Apr 2026). Under these conditions, “the RL policy maintains 100 % velocity-tracking success with MAE 4.0 cm/s (STD 7.1 cm/s) and perfect station keeping (4.5 cm range, 3.0 cm/s residual), whereas a tuned LQR baseline falls to 56 % success with MAE>23 cm/s under random orientation and fails station keeping entirely” (Wan et al., 27 Apr 2026).

The ablation results identify two ingredients as essential: “both friction randomization and IMU noise injection are essential: removing either drops tracking success to ~80 %” (Wan et al., 27 Apr 2026). This finding is consistent with the paper’s central thesis that the reality gap is dominated not by gross rigid-body mismatch but by frictional uncertainty and sensing imperfections.

On hardware, the paper reports “zero-shot Sim2Real across ceramic, carpet, slope and bump transitions” (Wan et al., 27 Apr 2026). For station keeping, the robot “drifts less than 5 cm on tile and 3 cm on a yoga mat, with residual speeds ~5 cm/s” (Wan et al., 27 Apr 2026). In “push-recovery trials (0.3 m lateral pushes),” the system “succeed[s] 7/7 times” (Wan et al., 27 Apr 2026), and “velocity tracking under human joystick control yields MAE 0.05 m/s” (Wan et al., 27 Apr 2026). The robot is also reported to “survive repeated falls and retain serviceability within seconds” (Wan et al., 27 Apr 2026).

These results sharply distinguish asRoBallet from prior dance-oriented robotic systems whose technical emphasis was not sim-to-real stabilization. For example, “Creative Robot Dance with Variational Encoder” focuses on real-time generation of “new dancing movements according to to the listened music” on a NAO humanoid (Augello et al., 2017), while “Breathless” emphasizes sinusoidal motifs, pose extraction with OpenPose, and live teach-mode interaction on a UR5e industrial arm (Cuan et al., 2024). asRoBallet instead uses expressive motion as a downstream capability built on robust underactuated locomotor control (Wan et al., 27 Apr 2026).

6. Position within robotic choreography and expressive robotics

asRoBallet belongs to a broader research landscape in which movement is treated as a computational, perceptual, and embodied medium. “Choreographic and Somatic Approaches for the Development of Expressive Robotic Systems” argues that “professional choreographers, dance practitioners, and movement analysts are critical to research in robotics,” and introduces frameworks based on Body, Effort, Space, and Shape, together with embodied exercises and rapid-prototyping tools for movement-centered design (LaViers et al., 2017). That line of work is concerned with how movement qualities are interpreted by human viewers and how movement vocabularies can be formalized for robot control (LaViers et al., 2017).

Other related work explores different technical slices of robotic dance and performance. “Creative Robot Dance with Variational Encoder” uses a VAE with “latent dimension $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$6” driven by music features such as “loudness” and “variance” to generate upper-body dance motion for a NAO humanoid (Augello et al., 2017). “Dance Teaching by a Robot” combines “adaptive impedance-based controller” design with “Progressive Teaching (PT)” scoring for partner-dance training, showing statistically significant gains in “comfort,” “peace of mind,” and “robot performance” at the $$M_{\mathrm{roll},i} = -c_r\,\dot\theta_{r,i}\,,\;i\in\{x,y\}$$7 level (Granados et al., 2018). “Breathless” presents an eight-hour duet between a human dancer and a UR5e using per-joint sinusoidal trajectories, OpenPose-based motion extraction, teach mode, and force-based cueing (Cuan et al., 2024).

Within this landscape, asRoBallet contributes a different axis of progress: not movement symbolism, audience interpretation, or pedagogical haptics, but the resolution of friction-sensitive underactuated dynamics sufficiently well that RL can be deployed “zero-shot” on a humanoid ballbot hardware (Wan et al., 27 Apr 2026). This suggests a possible convergence between expressive robotics and contact-rich mobile control: once stable and transferable balance is available, choreographic frameworks from expressive robotics (LaViers et al., 2017) and movement-generation systems (Augello et al., 2017) could, in principle, be layered atop a ballbot substrate.

A common misconception would be to treat asRoBallet primarily as a robot-performance project in the same sense as robot dance installations or theatrical systems. The paper’s actual emphasis is narrower and more technical: it is a Sim2Real control study centered on friction modeling, PPO training, and underactuated spherical dynamics, with expressive teleoperation and humanoid embodiment included as consequential platform features rather than the sole scientific endpoint (Wan et al., 27 Apr 2026).

7. Significance, limitations, and research implications

The principal significance claimed for asRoBallet is that it “closes the long-standing reality gap for underactuated spherical dynamics” by “explicitly simulating discrete roller mechanics, modeling coupled friction channels, and weaving domain-randomized RL with a consumer-grade iOS control loop” (Wan et al., 27 Apr 2026). The paper further states that the “subtractive-reconfigured hardware and all-in-one iOS interface democratize humanoid ballbots for future research in dynamic mobile manipulation and multi-agent coordination” (Wan et al., 27 Apr 2026).

Several limitations are implicit in the reported scope. The RL framework is described in terms of “two base policies—velocity tracking and station keeping” (Wan et al., 27 Apr 2026), rather than a unified whole-body manipulation policy. The strongest empirical claims concern balancing, velocity tracking, station keeping, and push recovery across surface transitions (Wan et al., 27 Apr 2026). This suggests that the work establishes a control foundation rather than a complete autonomous choreography system. A plausible implication is that future extensions would need to integrate the robust base policies with higher-level motion organization, embodied movement vocabularies, or interaction objectives of the kinds studied in expressive robotics (LaViers et al., 2017), dance generation (Augello et al., 2017), or performance systems (Cuan et al., 2024).

In that sense, asRoBallet can be understood as a control-centric foundation for expressive humanoid ballbots. Its novelty lies in treating wheel–sphere–ground tribology, actuator friction, observation noise, and embodied hardware redesign as inseparable components of a single Sim2Real pipeline (Wan et al., 27 Apr 2026). For researchers interested in underactuated mobile manipulation, physically expressive robotics, or low-cost high-agility platforms, the work is significant not because it replaces prior choreographic or artistic systems, but because it makes a previously fragile embodiment class available to RL-based deployment on real hardware (Wan et al., 27 Apr 2026).