Dr.-Ing. Raphael Deimel

Roboticist

raphael.deimel@tu-berlin.de

Current Interests / Past Projects

Human-Robot Interaction facilitated by Body Language

Phase-State Machines

Body language - communicating with body motion - is a neglected opportunity in human-robot interaction. Between humans, cues from movements and stance of the interaction partner(s) is used pervasively to negotiate interaction variants and to coordinate their temporal and spatial parameters, e.g. for object handover. The information from body motion is timely, hard to fake, but also continuous and volatile, which are extremely difficult to implement with our current robot's standard behavior generation architectures that are essentially based on finite state machines. I investigate a completely new approach on generating behavior that is able to adapt behavior smoothly and immediately and at any time, while also being interpretable as enacting a state graph. This hybrid continuous-discrete model called phase-state machine enables consistent and continuous modification of interaction state that is suited for interactive behavior in general and for communication using body language in particular.

Mixing diverse control goals using mechanical state distributions and Probabilistic Mechanical Primitives (ProMePs)

In any reasonably complex robot system, the need for blending or mixing between control goals arises. How can we do this in a principled manner even if control goals are diverse and even contradictive, such as position, velocity, force and stiffness? How to mix between deterministic PD-control goals and probabilistic goal descriptions such as Probabilistic Movement Primitives (ProMP)? And how can static, time-agnostic control goals and phase-based trajectories be reconciled if phase velocity is not predetermined? It seems like distributions over the full mechanical state of a robot, i.e. derivatives of both motion and effort, might just do those tricks. By stating an equivalence relationship between PD-control goals and desired distributions over the full mechanical state of a system we achieve a universal, consistent representation of control goals (mechanical state distributions) that has a clear intepretation and straightforward convertibility into control values for time-tested PD-control. At the same time, having the probabilistic description enables the mixing of arbitrary control goals in an information-theoretically principled way. Finally, mechanical state distributions are agnostic to how they are derived, providing an general API for mixing and blending disparate control strategies. A direct application of the above insights is the development of so-called Probabilistic Mechanical Primitives (ProMeP), which provide a method to learn and enact phase-parameterized and continuous control goals, as well as provide expectations about trajectories for both motions and forces. The latter can be used to e.g. guard the robot against applying forces or moving to locations that were not taught during learning and hence may be unsafe.

Light projection based robot head


We develop a device for robots to intuitively communicate intent, attention and other relevant social feedback during interaction scenarios in an intuitive manner, i.e. a robot head. Fluent and reactive behavior tie into a well designed appearance to create a consistent illusion to human interaction partners, In addition to the design and behavioral challenges, we also work on algorithms to optimize the optical system for arbitrary shapes of projection surfaces.

The PneuFlex actuators and toolkit

Online PneuFlex mold generator



I am working on bringing a mold generator online, which will enable you to request STL models of mold parts for your desired shape parameters. (currently on hold, contact me if interested)

PneuFlex Tutorial


This tutorial shows you how to build your own soft continuum actuators.

PneumaticBox


Today's commercial valve systems unfortunately are not built with robot system software integration in mind. The Pneumaticbox makes it easy to configure and run controllers in real-time with millisecond-precision, and to remote-control multi-channel pneumatic hardware such as the RBO Hand 2.

Soft Hands / Soft Manipulation

Soft Manipulation

Soft Manipulation is a project of the EU Horizon2020 programme, which aims to discover how robots can grasp and manipulate reliably with soft, compliant robot hands (2014-2019).

RBO Hand 1 and RBO Hand 2


The RBO Hands are an attempt to investigate the advantages and limits of completely soft hands. As the finger have no bones they cannot break, making the hand robust. At the same time, their shape adaptability makes grasping irregular or uncertain shapes as easy as inflating a balloon.

Soft hand Simulator based on SOFA


The soft hand simulator greatly simplifies development and evaluation of PneuFlex based robotic devices, such as soft hands. Based on the SOFA framework, it can configure and place actuators, "glue" them together and also run controllers to inflate/deflate the actuators and move the wrist around.
(please contact me if you are interested)

Automated Co-Design of Soft Hands / Morphological Computation

With soft robots, the distinction between control system and body becomes blurred in terms of who generates reactive behavior. As the body gains complexity, it also gains computational capacity. This effect is also called Morphological Computation. We investigate methods to quantify the contribution of body to generating useful behavior, and to use those measures for guiding automated design.
(please contact me if you are interested)

Publications

Download as bibtex file

Academic History

Distinctions