Legged Locomotion

Robust Locomotion in the Wild

We enable legged robots to confidently navigate complex, unstructured environments—muddy trails, snowy slopes, dense vegetation, and rushing streams—without any manual tuning, pre­mapping, or environment-specific adjustments. By training versatile control policies entirely in simulation with randomized terrains and physics parameters, then deploying them zero-shot on our ANYmal platform, we achieve “plug-and-play” locomotion - from multi-hour alpine ridge traverses to autonomous subterranean exploration - demonstrating robustness and energy efficiency across a wide range of real-world settings. Our approach blends proprioceptive and exteroceptive sensing through external page end-to-end learning architectures, enabling the robot to perceive and adapt to unseen obstacles and changing ground compliance on the fly. Beyond demonstrating state-of-the-art performance in the DARPA Subterranean Challenge and long-duration field trials, we’re advancing toward continuous, self-sufficient operation. Looking ahead, our research is aimed at extending these capabilities into even more extreme natural terrains—from dense forests on Earth to steep craters on the moon—pushing the boundaries of where autonomous robots can go and what they can achieve.

Contacts:
external page Filip Bjelonic
external page Clemens Schwarke
external page Chong Zhang
external page Takahiro Miki

Athletic Locomotion

Our research expands the frontiers of robotic agility, enabling legged robots to perform the kind of highly dynamic and acrobatic maneuvers seen in parkour. This capability is essential for robots to navigate cluttered disaster scenarios or rapidly traverse unstructured natural terrain where speed and versatility are critical. We develop end-to-end learning frameworks that empower both quadrupedal and humanoid robots to perform these maneuvers with speed, precision, and robustness. Key aspects of our work include training unified controllers by external page distilling specialized expert policies and developing advanced perception models that fuse sensory data for precise foot placement on external page challenging terrain. By including a clever foot design, our quadrupeds are also capable of advanced skills such as external page ladder climbing. Building on these capabilities, our research increasingly targets the complex domain of humanoid robots. Future work will explore how human motion data can guide policy learning, aiming to tackle the high dimensionality of humanoids and accelerate the acquisition of natural, human-like agility.

Contacts:
external page Clemens Schwarke
external page Junzhe He
external page Filip Bjelonic
external page Chenhao Li
external page Robert Baines

Low-Gravity Locomotion

The exploration of low-gravity bodies, such as the Moon or asteroids, calls for novel locomotion systems. Traditional wheeled rovers suffer from loss of traction in these environments. Legged robots, on the other hand, can exploit the lower gravity by adjusting their gait. At RSL, we perform research on learning-based locomotion controllers that are not only robust, but also adapted to their respective gravity environment. The emerging gaits range from crawling to running to leaping. Additionally, we set a special focus on power efficiency, as power and thermal budgets in space environments are much more restrictive than on Earth.

Contacts:
external page Philip Arm
external page Oliver Fischer
external page Filip Bjelonic

Effective Sim2Real Transfer

For simulation-based reinforcement learning to succeed in the real world, capturing accurate system dynamics is crucial. One effective strategy is the actuator network, which learns a model of the actuator response from real-world data, enabling physics simulators to replicate nuanced hardware behaviors such as delays, hysteresis, and saturation. This approach improves the fidelity of simulated training and allows policies to transfer more reliably to physical robots. Looking ahead, our research is focused on closing the sim-to-real gap even further—leveraging hybrid modeling, uncertainty-aware learning, and self-supervised adaptation to build models that generalize across hardware platforms, wear states, and environmental conditions. By combining high-fidelity simulation with intelligent adaptation strategies, we aim to enable robots that learn faster, deploy sooner, and perform robustly in environments never seen before. Our current research will be published soon, stay tuned.

Contact:
external page Filip Bjelonic

Learning from Demonstrations

Learning from demonstrations (LfD) leverages offline state-based reference data to guide policy learning in reinforcement learning (RL). At the Robotic Systems Lab, our research focuses on the use of expert demonstrations—often collected from motion capture, teleoperation, or high-fidelity simulations—not merely as a source of imitation, but as a structured learning signal that drives efficient skill acquisition in high-dimensional control problems. Demonstrations serve as inductive priors that regularize policy learning, biasing exploration toward semantically meaningful regions of the state-action space. Especially in legged locomotion and other high-DoF systems, this guidance is crucial: as the dimensionality of the problem grows, manual reward shaping or curriculum design becomes less viable. In contrast, leveraging demonstrations allows for more scalable and data-efficient learning.

Contact:
external page Chenhao Li

Online Learning

Robotic systems have achieved remarkable advancements in recent years, driven by progress in reinforcement learning (RL) and control theory. A prevalent limitation in many approaches is the lack of adaptation and learning once the policy is deployed on the real system. This results in underutilization of the valuable data generated during real-world interactions. Robotic systems operating in dynamic and uncertain environments require the ability to continually adapt their behavior to new conditions. The inability to exploit real-world experience for further learning restricts the system’s robustness and limits its ability to handle evolving scenarios effectively. Truly intelligent robotic systems should operate efficiently and reliably using limited data, adapting to real-world conditions in a scalable manner.

Contact:
external page Chenhao Li