ANYmal C Robot Platform

• ANYmal C is a fully electric, 12-degree-of-freedom quadrupedal platform that integrates advanced actuation and sensor fusion for agile performance in complex terrains.
• The system features high-torque series-elastic actuators with embedded PD/impedance control, ensuring precise, rapid motion and effective sim-to-real transfer.
• Leveraging reinforcement learning alongside model-based planning, ANYmal C achieves efficient, high-speed locomotion and robust navigation in unstructured environments.

ANYmal C is a fully electric, 12-degree-of-freedom quadrupedal robot platform developed by ETH Zürich and deployed by the company ANYbotics. It is designed for dynamic locomotion, robust mobility in unstructured environments, and rapid research prototyping at the intersection of reinforcement learning, optimal control, and field robotics. The platform supports state-of-the-art sensing, high-rate actuation, and a broad range of locomotion, manipulation, and navigation capabilities. ANYmal C provides a reference architecture for academic and industrial deployment of legged robots.

1. Physical Architecture and Hardware Platform

ANYmal C features a rigid carbon-fiber main body with four identical legs mounted at the chassis corners. Each limb provides three actively torque-controlled joints (hip ab/adduction, hip flexion/extension, knee flexion/extension) implemented via high-torque series-elastic actuators (SEA) with integrated harmonic-drive gearboxes. Key specifications across multiple studies are as follows:

Parameter	Value (typical range)	Reference
Mass (with batteries)	30–55 kg	(Scheidemann et al., 1 Oct 2024, Papatheodorou et al., 2023)
Degrees of Freedom	12 (3 per leg)	(Scheidemann et al., 1 Oct 2024, Geisert et al., 2019)
Joint torque limits	±40–80 Nm	(Hwangbo et al., 2019, Gangapurwala et al., 2020)
Joint speed limits	12 rad/s	(Jadoon et al., 23 Feb 2025, Hwangbo et al., 2019)
Footprint (L × W × H)	~0.6 m × 0.4 m × 0.6 m	(Scheidemann et al., 1 Oct 2024)
Payload capacity	~10 kg	(Scheidemann et al., 1 Oct 2024)

Onboard computing architectures combine multicore Intel CPUs (Core i7 class), high-performance GPUs (e.g., NVIDIA Jetson Orin), high-speed buses (Ethernet/CAN), and redundant power supplies (Scheidemann et al., 1 Oct 2024). Sensing modalities include proprioceptive (IMU, joint encoders, torque sensing), exteroceptive (multiple synchronized RGB-D cameras, 3D LiDAR, custom RF AoA beacons), and optional contact and force sensors (Scheidemann et al., 1 Oct 2024, Shi et al., 2020).

2. Kinematic, Dynamic, and Actuator Modeling

The floating-base kinematics are formalized as a 6-dof body with 12 actuated joints. Forward kinematics use Denavit–Hartenberg conventions specific to ANYmal’s layout (Geisert et al., 2019), with each limb designed for nearly spherical hip rotation and extensive lateral reach. The full system dynamics are typically expressed as

$$M(q)\ddot{q} + C(q, \dot{q})\dot{q} + g(q) + J_c(q)^T \lambda = S^T \tau$$

where $q$ concatenates floating-base and joint angles; $M$, $C$, $g$ are inertia, Coriolis, and gravity; $J_c$ maps contact forces $\lambda$; $S$ selects actuated joints; $\tau$ are actuator torques (Geisert et al., 2019).

Low-level joint actuation is managed by embedded PD/impedance control, with typical gains in the range $k_p = 50$–$80\,\text{Nm/rad}$, $k_d = 0.1$–$2\,\text{Nm}\,\text{s/rad}$ (Hwangbo et al., 2019, Gangapurwala et al., 2022). Recent work demonstrates that data-driven actuator networks, fit to real motor data, significantly outperform analytical models for robust sim-to-real transfer and energy efficiency (Jadoon et al., 23 Feb 2025, Hwangbo et al., 2019). Joint position, torque, and velocity limits are strictly enforced both in simulation and on hardware.

3. Control, Planning, and Locomotion Algorithms

ANYmal C supports a spectrum of motion generation approaches from model-based trajectory optimization to deep RL. Key approaches include:

• Model-based contact/trajectory planning: Reachability-based kinodynamic planning (Geisert et al., 2019) generates collision-free body paths and greedy, acyclic contact sequences via precomputed octrees of valid foothold configurations. This approach can plan 50-step whole-body motions in under 10 s for moderately complex environments.
• Full-centroidal dynamics optimization: Efficient OCP formulations employ implicit inverse kinematics and analytic centroidal dynamics to synthesize high-acceleration maneuvers such as squat and rotational jumps while respecting hardware torque limits (Papatheodorou et al., 2023).
• Reinforcement learning (RL):
  • TRPO and PPO-based policies trained in simulation deliver robust, high-speed trotting, energy-efficient gait transitions, and rapid recovery from falls (Hwangbo et al., 2019, Jadoon et al., 23 Feb 2025, Gangapurwala et al., 2022).
  • RLOC integrates terrain-aware RL footstep planning with whole-body optimal control and learned domain-adaptive torque trackers for zero-shot transfer across different ANYmal generations (Gangapurwala et al., 2020).
  • GCPO (Guided Constrained Policy Optimization) embeds safety, kinodynamic, and stability constraints directly into the RL update pipeline, converging 10× faster than unconstrained PPO and achieving near-zero hardware constraint violations (Gangapurwala et al., 2020).

A typical high-rate control stack closes local torque loops at 400 Hz, while higher-level planners and learned policies run at 10–100 Hz depending on the architecture and task (Scheidemann et al., 1 Oct 2024, Gangapurwala et al., 2022).

4. Perception, Mapping, and Sensor Fusion

ANYmal C’s field autonomy relies on rich sensor fusion spanning RF, vision, and 3D point cloud modalities. Notable integrated systems include:

• RF AoA beacon tracking: Custom-built beacon and circular phased-array receiver architectures provide 3° mean angular accuracy under nominal indoor conditions and up to 7° in challenging multipath scenarios at 5 Hz update (Scheidemann et al., 1 Oct 2024).
• LiDAR and RGB-D fusion: Four Intel Realsense D265 modules and a Velodyne VLP-16 LiDAR (360° × ±15° FOV) are rigidly mounted and merged at rates up to 30 Hz (cameras) and 10 Hz (LiDAR), with calibration via hand–eye procedures (Scheidemann et al., 1 Oct 2024).
• Leader estimation and obstacle avoidance: Human detection is performed by YOLOv8 networks on RGB, associated with AoA estimates and LiDAR clusters. Leader state is fused via an EKF with data association and robust outlier rejection (Scheidemann et al., 1 Oct 2024).
• Riemannian Motion Policies (RMPs): Reactive RMP planners encode goal attraction, static/dynamic obstacle repulsion, and yaw alignment as metric-weighted flows in $\mathrm{SE}(2)$, merged for real-time kinodynamic reference updates (Scheidemann et al., 1 Oct 2024).

This perception pipeline supports robust leader following and navigation in crowded, cluttered, and dynamic environments at up to 0.4–0.5 m/s, with typical tracking latency below 120 ms (Scheidemann et al., 1 Oct 2024).

5. Sim-to-Real Transfer and Experimental Performance

Empirical studies consistently demonstrate high-fidelity sim-to-real transfer:

• RL-driven controllers learned in RaiSim and other fast rigid-body/contact simulators, with heavy domain/inertia/noise randomization, transfer directly to ANYmal C hardware without reward retuning or manual parameter adjustment (Jadoon et al., 23 Feb 2025, Hwangbo et al., 2019, Gangapurwala et al., 2020, Gangapurwala et al., 2022).
• Forward velocity commands up to 1.6 m/s are reliably tracked with high-energy efficiency and robust gait recovery on flat and moderately rough terrain (Hwangbo et al., 2019, Jadoon et al., 23 Feb 2025).
• GCPO policies deliver RMS velocity tracking errors of ≈0.05–0.1 m/s up to 1 m/s on real hardware, outperforming model-based gait controllers while strictly respecting joint and contact constraints (Gangapurwala et al., 2020).
• The RLOC framework achieves >80% success rates over challenging stair/brick/wave terrains and can reject external perturbations up to 400 N (Gangapurwala et al., 2020).
• Low-frequency learned motion controllers (down to 8 Hz) prove robust to actuation latency up to 90 ms and maintain successful sim-to-real transfer without explicit actuator model learning or dynamics randomization (Gangapurwala et al., 2022).
• Leader-following and manipulation tasks demonstrate multi-modal fusion and agile, contact-rich interaction capabilities absent from earlier quadrupedal platforms (Scheidemann et al., 1 Oct 2024, Shi et al., 2020).

6. Advanced Behaviors: Dexterous Manipulation and Industrial Use

ANYmal C extends beyond locomotion to non-prehensile limb-based manipulation and industrial navigation testing:

• Circus ANYmal: Demonstrates robust dynamic ball manipulation (max rotation ≈15°/s, 95% recovery after perturbation) using only proprioceptive feedback, RL, and aggressive domain/coupling randomization—no dedicated manipulators or tactile sensors (Shi et al., 2020).
• Industrial simulation-based testing: ANYmal C is validated in extensive simulation-based verification (Surrealist), autogenerating adversarial obstacle courses and benchmarking proprietary navigation algorithms, improving success rates from 40.3% to 71.2% in pilot phases and driving rapid iterative improvement during industrial evaluation (Khatiri et al., 10 Oct 2025).
• Performance envelopes: The platform’s hardware and learning-based software stack unlock flight-phase trotting, high-acceleration acrobatic maneuvers (linear/rotational jumps), and high-throughput real-world deployment in inspection, rescue, and assistive roles (Papatheodorou et al., 2023, Scheidemann et al., 1 Oct 2024).

7. Lessons, Limitations, and Future Directions

Empirical findings emphasize the following:

• Multi-modal sensor fusion and explicit decoupling of static vs. dynamic obstacles enhance robustness and reduce mapping noise (Scheidemann et al., 1 Oct 2024).
• Learned actuator models and aggressive domain randomization are essential for closing the sim-to-real gap in energy-efficient, high-speed control (Hwangbo et al., 2019, Jadoon et al., 23 Feb 2025).
• RL with embedded constraints (GCPO) achieves superior hardware safety and convergence over unconstrained approaches (Gangapurwala et al., 2020).
• Low-frequency motion policies and simple impedance tracking dramatically improve latency robustness and sim-to-real faithfulness (Gangapurwala et al., 2022).
• Task-space full-centroidal planning with analytic derivatives enables real-time high-acceleration motion synthesis, but ignores internal limb momentum effects (Papatheodorou et al., 2023).

Limitations include the need for further integration of whole-body dynamic constraints in reactive planners (Scheidemann et al., 1 Oct 2024), complete sim-to-real transfer for nontrivial terrain and manipulation tasks, and joint optimization of perception and control (Shi et al., 2020). Future research aims at incorporating whole-body SE(3) RMPs, learned cost/metric weights, onboard vision for manipulation, automated scenario generation from real-world logs, and MPC integration in dynamic planning.

ANYmal C’s modularity and documented performance across RL, model-based control, and real-world deployment situate it as a central research and industrial platform for legged robotics (Scheidemann et al., 1 Oct 2024, Jadoon et al., 23 Feb 2025, Gangapurwala et al., 2020, Khatiri et al., 10 Oct 2025).