Postdoctoral Positions in Robot Learning and Soft, Musculoskeletal, and Biohybrid Robotics

100%, Zurich, fixed-term

print Drucken

The Soft Robotics Lab within the Institute of Robotics and Intelligent Systems at ETH Zurich is inviting applications for several open postdoctoral positions. Our lab's goal is to build, model, and control robots in a fundamentally different way, so that they become more flexible, dexterous, capable, and adapt better to their environment. We work along four directions: soft and musculoskeletal robotics, biohybrid living systems, dexterous manipulation and robot learning, and simulation for embodied AI.

We are looking for exceptional candidates in any of them. This round we especially want two profiles: people who design and build the robots, and people who make policies run on them.

We do not hire against a narrow project description. We hire people who will define their own. Tell us which of our directions you want to push, and why you are the person to push it.

Project background

Today's robots are mostly rigid, fragile, and a world apart from the agility and resilience of biological bodies. Our bet is that the next generation of robots will be soft, musculoskeletal, and in part alive. They will be built to make contact with the real world rather than to avoid it. We pursue this across four directions, and a strong candidate will find a home in one of them and borrow from the others.

We build bodies from compliant structures, bones, joints, and tendon‑like actuation. Our electrohydraulic musculoskeletal leg jumps, moves fast, and adapts to terrain at roughly 1.2 % of the energy a motor‑driven leg needs (Nature Communications, 2024). Our low‑voltage HASEL actuators run near 1100 V, are safe to touch, and work untethered and underwater (Science Advances, 2024). We recently extended these muscles to full antagonistic motion ranges (ICRA 2025) and to a sensorless, inherently compliant anthropomorphic hand driven entirely by electrohydraulic actuation (IROS 2026).

We grow engineered muscle and use it to actuate machines. We bioprinted multicellular muscle‑tendon units that transmit force along a real musculoskeletal path (Science Advances, 2025), embedded sensors directly into muscle for closed‑loop control of proprioceptive biohybrid robots (Advanced Intelligent Systems, 2025), and established functional volumetric bioprinting with xolography (Advanced Materials, 2026). Co‑optimized volumetric muscle designs for large dynamic deformations are in press at Nature Communications (Balciunaite et al.,

2026). The same fabrication line reaches clinical work: with University Hospital Zurich we printed implantable reinforced cardiac tissue patches (Advanced Materials, 2025).

We build hands and the policies that run them. One of our initial hand designs is now commercialized through our spin‑off Mimic Robotics. ORCA is our open‑source, reliable, and cost‑effective anthropomorphic hand for uninterrupted dexterous task learning (IROS 2025). On that hardware we work on imitation learning and diffusion policies, cross‑embodiment skill transfer through latent action diffusion (ICRA 2026), sample‑efficient reinforcement learning and policy fine‑tuning directly on the real robot, vision‑language‑action models for contact‑rich tasks, and tactile representation learning on our high‑resolution sensorized skin (ICRA 2024).

We also build controllable dexterous world models for training and evaluation, and a benchmark of dexterity for anthropomorphic hands. Whichever side you come from, the offer is the same: the hand, the skin, the simulator, and the people who designed all three sit in one room. Build a mechanism here and someone will have a policy running on it within weeks.

Build a policy here and you can change the mechanism when the mechanism is what is wrong.

Building these robots requires tools that did not exist. Vision‑Controlled Jetting prints rigid skeletons, soft tissue, tendons, and sensors in one pass, including a full musculoskeletal hand and forearm (Nature, 2023). We close the sim‑to‑real gap with learned residual physics…