Autonomy is one of the most significant scientific milestones challenging the next generation of roboticists. Not only will research in autonomy bring a host of new robotics applications to market, but it will also support new advances in our understanding of human beings. In this presentation, I discuss an experimental architecture for studying autonomy. In particular, we propose a landscape of attractor-based architecture that supports a unique state representation (called aspects) and mechanisms for learning from direct exploration—generating artifacts including a library of re-useable skills and an explicit representation of the environment in terms of the transition dynamics. Finally, I will present some of our newest work on approximate solutions to belief-state MDPs that makes use of past experience to actively gather state information and to guarantee performance in the course of problem solving in partially observable systems.

Extrapolating these results, we are looking forward to technological artifacts that support a new method for determining non-uniform and time-varying sensitivity to uncertainty and architectures for tolerating uncertainty where possible and actively suppressing aspects of uncertainty that could lead to unacceptable outcomes. I will highlight experimental results using our laboratories mobile manipulator (uBot) that confirms parts of our hypothesis---that a representation consisting of transitions between patterns of co-activated controllers is an effective encoding of environmental control affordances that supports active perception and task-level control simultaneously. I will demonstrate experimental results that: continuously monitor uncertainty; recognize and exploit probabilistic roadmaps in the state transition function; and eliminate unrecoverable errors in autonomous robot systems.

In the design of autonomous cars, the focus typically is on optical sensors, be it regular cameras, or Lidar, or radar and usually a combination of all of these. In the talk I will discuss some of the work that also we did along those lines. That work includes semantic scene segmentation, instance segmentation, lane detection, depth map reconstruction, tracking, etc. But then there also is work needed that goes beyond the interpretation of such sensor input. Example topics are:
- the use of maps that - as a counterpart to the cases where people know the environment in which they drive - add strong knowledge about what is to come,
- the ability to talk to your car, thus offering a user-friendly and intuitive way to give indications about desired driving style or how to change the planned route, and
- the addition of aural input, as human drivers also constantly listen to the sounds made by their car (e.g. type of road) and other traffic agents around them.

Agriculture has benefited from autonomous robots for a considerable time. Since the 1980’s researchers and companies around the world have been developing new equipments to assist farmers aiming to increase food production and reduce costs. When it comes to the production chain, the scope of their application cover soil preparation, pre-harvesting, harvesting, post-harvesting, distribution, etc. Mainly concentrated in metropolitan areas, the world population is growing, which makes these robots extremely necessary. Not to mention the climate change and environmental issues.

Contrary to what is widely believed, developing autonomous robots for agriculture has proved to be complex and time consuming. Field robots have to cope with temperature, brightness and humidity variations. The presence of unknown obstacles, such as mud, holes, stones, fences, soil erosion, animals, machines, people and other crops is also challenging.

For this talk, I have prepared a short historic overview and the challenges we have been facing. USP-EESC-LabRoM in cooperation with partners in Brazil and abroad are carrying out some projects I will present you, hoping we have a glance of the next steps we will take.