Hello, I'm Oliver Hausdörfer

Tutorial marking ten years of CDC Learning-Based Control sessions, covering stability and safety foundations, dynamics learning, data-driven control, and RL/IL/BO, with open-source examples.

An introductory guide, marking ten years of dedicated conference sessions, on how robots and other machines can learn from data and experience while staying stable and safe, with open-source code.

Multi-view latent action pretraining improves 6D action decodability by 17% over single-view, outperforming 100× more single-view data, and lifts hardware success from 21% to 46%. We demonstrate zero-shot behavior transfer from human videos.

Robots can learn skills from videos of other robots or human hands. Filming each video from two cameras instead of one shows movements a single camera misses, more than doubling real-robot success. It shows which data are most useful for robot learning.

55 scene-aligned 4D human motion sequence dataset from egocentric wearables for training perceptive humanoid locomotion policies on challenging terrain.

55 recordings of people climbing, crawling and balancing over obstacles, captured with wearable sensors and lined up with a 3D scan of the environment, to teach humanoids to cross rough terrain.

Autonomous data collection in the contact-rich segment yields 3–5× more data and better data coverage than teleoperation. Combining this data with offline RL and motion planning for free-space motion yields 96% success, vs. 37% for end-to-end learning from teleoperation.

The robot learns the challenging part of the task on its own, while off-the-shelf tools handle the rest. This reliably solves precise assembly tasks, with over 95% success. It even works well for new scenes.

Functionally decomposing locomotion into task reward, constraints, energy minimization and perception yields terrain-adaptive gaits without explicit gait priors, cutting cost of transport by 56% and limit violations by 96%, with zero-shot transfer to a Go2.

Splitting a robot dog’s training into separate jobs (follow commands, stay within its physical limits, save energy, see the terrain ahead) lets it find an efficient trot by itself, without hand-engineered specifications.

C++ Controllers, Python Interfaces, and Gymnasium environments for teleoperation and deploying learning-based policies on manipulators.

A software that lets different robot arms smoothly follow commands from AI models. It also provides tool to remote-control the arm and train those models. CRISP enables AI training and deployment on manipulators.

A single gait cycle of expert demonstration significantly improves achieved reward and visual appearance of the learned gait in locomotion tasks - even for transfer tasks.

An LLM can generate one short walking example, which helps a legged robot learn to walk better and more naturally — even in situations the example never covered, like steering or rough ground. This allows non-experts to easily program walking robots.

Inter-GPU communication compression for Large Language Models leads to up to 2x reduction of time-to-first-token with negligible model performance degradation.

Big AI models are split across multiple GPUs that constantly exchange data. By compressing this data 3.5–4.5x, we make them start answering up to twice as fast, with almost no loss in quality.

In a simulated lamprey robot without direct inter-segmental coupling, stretch feedback learned via deep RL coordinates swimming. Local, caudally directed, stretch-rate information proves key, and weight-shared local networks learn 2–3× faster and more consistently.

We use robots to investigate how animals like the eel-like lamprey move. With biologically inspired feedback, where each body section senses only how its neighbors bend, the robot learns faster and more reliably.

3D-printed chip for mechanoactivation at single-cell level to study intracellular calcium signaling and translocation.

A 3D-printed chip that squeezes many stem cells at once, each tucked into its own tiny gel bead, so we can watch under the microscope how single cells sense and respond to pressure. This could help make stem-cell therapies, such as repairing bone or cartilage, more effective.