A list of SRL project websites.
Preprint (arXiv), 2026
We previously provided a reinforcement learning policy for underwater tendon-driven swimmers, which we now characterize on a new fish hardware design, and design an experimental setup for real-time deployment.
ETH Focus Project, 2025/26
Robotic systems for space-station cargo, inventory, and maintenance tasks, freeing astronauts to focus on scientific research. The HELIOS humanoid prototype uses four arms to move, stabilize, and manipulate in microgravity.
Preprint (arXiv), 2026
An untethered, off-the-shelf anthropomorphic hand learns in simulation to crawl on its fingertips across 14 indoor and outdoor surfaces, recover from falls, press keys, and push objects while supporting its own weight.
Preprint (arXiv), 2026
Using only the CPU on a laptop, we propose a controller that can write with a real pen in hand on paper, after just 18 s of initialization, without an analytic hand-object kinematic/contact model, simulation training, or precollected task demonstrations. This showcases an alternative to compute- and data-heavy approaches such as RL and IL for achieving dexterous manipulation through computationally simple and data-efficient algorithms.
Preprint (arXiv), 2026
A systematically modernized offline reinforcement learning actor-critic with an exact-likelihood flow policy, categorical residual critics, staged training, and test-time action improvement.
Preprint (arXiv), 2026
EgoVerse is a living, consortium-scale egocentric dataset capturing diverse human demonstrations across tasks, scenes, and operators worldwide, purpose-built for human-to-robot transfer across embodiments.
Preprint (arXiv), 2026
POMDAR is a taxonomy-grounded dexterity benchmark for anthropomorphic hands, measuring task throughput across grasping and manipulation configurations in real and simulated settings.
IEEE/RSJ IROS, 2026
A smooth, low-dimensional basis-function embedding for jointly optimizing soft robot shape, material distribution, and actuation through black-box simulation.
IEEE ICRA, 2026
Digital twins for tendon-driven underwater robots, validating simple models with real-world swimming behavior.
Preprint (arXiv), 2026
Sample-efficient real-world dexterous policy fine-tuning with action-chunked critics and normalizing flows.
IEEE RoboSoft, 2026
A modular, high-fidelity simulation framework for soft robots focused on contact-rich manipulation and sim-to-real transfer.
IEEE ICRA, 2026
Latent action diffusion for cross-embodiment skill transfer and single-policy multi-robot manipulation control.
IEEE Transactions on Neural Networks and Learning Systems, 2025
Extensive real-world validation of the DeSKO robust learning-control framework on a physical SoPrA soft arm, with integral feedback enabling zero-shot pick-and-place of unseen objects and payloads.
IEEE ICRA, 2025
A HASEL-clutch artificial muscle system that lets non-stretchable electrohydraulic actuators work in antagonistic joints without sacrificing motion range to tendon slack.
Preprint (arXiv), 2025
A diffusion-based model recipe and data collection protocol for a 16-DoF tendon-driven dexterous hand, achieving up to 93.3% out-of-distribution success rates with emergent self-correcting behaviors.
Preprint (arXiv), 2025
MAPLE learns dexterous manipulation priors from large-scale egocentric videos, producing vision features that enable efficient policy learning and strong zero-shot generalization for dexterous robotic manipulation.
Science Advances, 2024
HALVE actuators lower electrohydraulic artificial-muscle operation to around 1100 V while remaining safe to touch, waterproof, and self-clearing, enabling untethered gripper and soft robotic fish demonstrations.
IEEE RoboSoft, 2024
A biomimetic hand and forearm 3D-printed in a single process—rigid skeleton, soft joint capsules, tendons, and touch sensors—actuated by 22 independently controlled pneumatic artificial muscles with human-like dexterity.
Nature Communications, 2024
A bio-inspired robotic leg driven by pairs of antagonistic electrohydraulic (HASEL) artificial muscles, capable of jumping, fast movement, and automatic terrain adaptation while consuming only 1.2% of the energy of a conventional motor-driven leg.
Preprint (arXiv), 2024
PokeFlex is a real-world multimodal dataset of 18 deformable objects—including 3D textured meshes, point clouds, RGB images, and depth maps—enabling online mesh reconstruction and sim-to-real transfer for deformable object manipulation.
Preprint (arXiv), 2024
Vector-quantized action chunk embeddings for efficient policy search in dexterous manipulation with MPC and reinforcement learning.
IEEE ICRA, 2024
Sensorized soft skin for high-resolution tactile perception in soft robotics.
IEEE RA-L, 2024
Hybrid sim-to-real modeling for soft robots that augments analytical simulation with learned residual physics.
IEEE IROS, 2024
Open source Robotic Hand.
CoRL, 2024
An open-source platform for autonomous aerial manipulation using only onboard perception, enabling independent grasping in diverse indoor and outdoor settings with a released ROS 2 software stack and custom hardware designs.
IEEE ICRA, 2023
Aquarium is a fully differentiable 2D FSI solver using implicit Navier-Stokes integration and immersed-boundary coupling, providing analytical gradients for gradient-based gait and shape co-optimization of underwater robots.
IEEE-RAS Humanoids, 2023
Adaptive dexterous manipulation for ball rolling tasks with soft robotic systems.
Nature, 2023
Vision-driven control and optimization for ink-jet based manufacturing.
ICLR, 2022
A deep stochastic Koopman operator that models nonlinear uncertainty as a distribution of observables, enabling efficient robust MPC with a closed-loop stability guarantee across soft-robotic, legged, and biological benchmarks.
IEEE/RSJ IROS, 2022
A quadcopter with a soft Fin Ray gripper for fast, robust aerial grasping, achieving 83% efficacy across four object geometries at ~1 m/s while supporting up to 4× the payload of prior works.