Sim-to-Real Transfer for Muscle-Actuated Robots via Generalized Actuator Networks
| Authors | Jan Schneider et al. |
| Year | 2026 |
| Field | AI / ML |
| arXiv | 2604.09487 |
| Download | |
| Categories | cs.RO, cs.LG |
Abstract
Tendon drives paired with soft muscle actuation enable faster and safer robots while potentially accelerating skill acquisition. Still, these systems are rarely used in practice due to inherent nonlinearities, friction, and hysteresis, which complicate modeling and control. So far, these challenges have hindered policy transfer from simulation to real systems. To bridge this gap, we propose a sim-to-real pipeline that learns a neural network model of this complex actuation and leverages established rigid body simulation for the arm dynamics and interactions with the environment. Our method, called Generalized Actuator Network (GeAN), enables actuation model identification across a wide range of robots by learning directly from joint position trajectories rather than requiring torque sensors. Using GeAN on PAMY2, a tendon-driven robot powered by pneumatic artificial muscles, we successfully deploy precise goal-reaching and dynamic ball-in-a-cup policies trained entirely in simulation. To the best of our knowledge, this result constitutes the first successful sim-to-real transfer for a four-degrees-of-freedom muscle-actuated robot arm.
Engineering Breakdown
Plain English
This paper addresses a critical gap in robotics: soft robots with tendon-driven actuation are faster and safer but extremely hard to control because their actuators exhibit nonlinearities, friction, and hysteresis that break simulation-to-reality transfer. The authors propose Generalized Actuator Network (GeAN), a neural network model that learns the complex dynamics of these actuators directly from joint position trajectories—no torque sensors needed—and pairs it with standard rigid body simulation for the arm and environment. This enables robots to acquire manipulation skills in simulation and successfully deploy them on real hardware, solving a longstanding problem that has kept these actuators out of practical use despite their advantages.
Core Technical Contribution
GeAN's core novelty is learning actuator models from trajectory data alone rather than requiring instrumented sensors or hand-crafted dynamics equations. Instead of modeling torque explicitly, the method learns a neural network that predicts how joint positions evolve given motor commands, automatically capturing nonlinearities and hysteresis without explicit parameterization. This is paired with a standard rigid body dynamics engine (like PyBullet or MuJoCo), creating a hybrid pipeline where learned actuator models feed into conventional physics simulation. The key insight is that you don't need to fully understand the actuator—you just need to predict its input-output behavior accurately enough for policy learning and transfer.
How It Works
The pipeline works in three stages: (1) collect trajectories from the real robot executing random or exploratory motor commands, recording joint positions and motor inputs; (2) train a neural network to map motor commands and current state to predicted joint accelerations or positions, effectively learning the actuator forward model; (3) embed this learned network into a physics simulator alongside rigid body dynamics and task rewards, then train RL policies in simulation that transfer directly to hardware. The learned actuator network acts as a plug-in component—it replaces the analytical actuator model that would normally be hand-derived or identified through expensive parameter optimization. During real deployment, the policy outputs motor commands, GeAN predicts the resulting joint motion, and the rigid body simulator propagates this into task-space forces and interactions with the environment.
Production Impact
For robotics teams, this eliminates the need for expensive torque/force sensors and manual system identification, reducing hardware cost and complexity. Policies trained in simulation transfer directly to real robots, dramatically cutting real-world training time and wear on hardware—critical for expensive manipulators. The main production trade-off is data collection overhead: you need representative trajectory data to train GeAN accurately, which means running exploration on real hardware upfront. Compute cost is modest (network inference is fast), but integration requires setting up a simulator-based policy training pipeline and ensuring your simulation's rigid body dynamics are reasonably accurate. Teams using soft or elastic actuation—collaborative robots, prosthetics, legged systems—benefit most; teams with stiff industrial arms may not justify the added complexity.
Limitations and When Not to Use This
GeAN assumes that learning actuator dynamics from trajectory data generalizes to the same actuator under different loads, speeds, and contact forces—this may not hold for extreme regimes or significantly different tasks. The method still requires a physics simulator with accurate kinematic/dynamic parameters; garbage simulation means garbage policies, even with perfect actuator models. The paper doesn't address how to handle actuator failure modes (jamming, cable slack, motor saturation) or online adaptation when hardware degrades over time. It also likely requires more exploration data than existing methods if actuators are highly nonlinear or hysteretic, and the approach assumes deterministic actuator behavior—true stochasticity (e.g., from cable noise or compliance variation) isn't explicitly modeled.
Research Context
This builds on a decade of sim-to-real work in robotics (starting with domain randomization and adversarial training) but specifically targets the actuator gap—prior work assumed you could model the motor faithfully or used expensive system identification. It relates to broader neural surrogate modeling literature (replacing simulators or components with learned approximations) but applies it to the actuator layer rather than full dynamics. The work fits into the soft robotics renaissance, where researchers are discovering that compliance enables faster, safer interaction but control is much harder. It opens up a research direction: learning modular actuator models that transfer across robot morphologies, potentially enabling actuator-agnostic policy libraries.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::