04. – 09. Oktober 2015, Dagstuhl Seminar 15411
Multimodal Manipulation Under Uncertainty
1 / 4 >
Auskunft zu diesem Dagstuhl Seminar erteilt
While robots have been used for decades to perform highly specialized tasks in engineered environments, robotic manipulation is still crude and clumsy in settings not specifically designed for robots. There is a huge gap between human and robot capabilities, including actuation, perception, and reasoning. However, recent developments such as low-cost manipulators and sensing technologies place the field in a good position to make progress on robot manipulation in unstructured environments. Various techniques are emerging for computing or inferring grasp configurations based on object identity, shape, or appearance, using simple grippers and robot hands.
Beyond grasping, a key ingredient of sophisticated manipulation is the management of state information and its uncertainty. One approach to handling uncertainty is to develop grasping and manipulation skills that are robust to environmental variation. Another approach is to develop methods of interacting with the environment in order to gain task-relevant information, for example, by touching, pushing, changing viewpoint, etc. Managing state information and uncertainty will require a tight combination of perception and planning. When the sensor evidence is unambiguous, the robot needs to be able to recognize that and perform the task accurately and efficiently. When greater uncertainty is present, the robot needs to adjust its actions so that they will succeed in the worst case or it needs to gain additional information in order to improve its situation. Different sensing modalities as well as world models can often be combined to good effect due to their complementary properties.
This seminar discussed research questions and agendas in order to accelerate progress towards robust manipulation under uncertainty, including topics such as the following:
- Is there a master algorithm or are there infinitely many algorithms that solve specialized problems? Can we decompose multimodal manipulation under uncertainty into I/O boxes? If so, what would these be?
- Do we prefer rare-feedback / strong-model or frequent-feedback / weak-model approaches? Is there a sweet spot in between? Is this the way to think about underactuated hands?
- What are useful perceptual representations for manipulation? What should be the relationship between perception and action? What kind of perception is required for reactive systems, planning systems, etc.?
- How do we do deformable-object manipulation? What planning methods, what types of models are appropriate?
- How should we be benchmarking manipulation? What kind of objects; what kind of tasks should be used?
- How should humans and robots collaborate on manipulation tasks? This question includes humans collaborating with autonomous robots as well as partially-autonomous robots acting under human command.
In the area of perception, we concluded that the design of representations remains a central issue. While it would be beneficial to develop representations that encompass multiple levels of abstraction in a coherent fashion, it is also clear that specific visual tasks suggest distinct visual representations.
How useful or limiting is the engineering approach of decomposing functionality into separate modules? Although this question was heavily debated, the majority view among seminar participants was that modules are useful to keep design complexity manageable for humans, and to keep the event horizon manageable for planning systems. It seems that to build more flexible and powerful systems, modules will need to be more strongly interconnected than they typically are these days. Fundamental challenges lie in the specification of each module and of their interconnections. There is a lot of room for creative innovation in this area.
Benchmarking questions were discussed chiefly in the context of the YCB Object Set. Specific benchmarks were suggested and discussed, covering perception and planning in the context of autonomous manipulation.
Creative Commons BY 3.0 Unported license
Jan Peters and Justus Piater and Robert Platt and Siddhartha Srinivasa
- Artificial Intelligence / Robotics
- Computer vision
- Range sensing
- Tactile sensing