We present DemoStart, a novel auto-curriculum reinforcement learning method capable of learning complex manipulation behaviors on an arm equipped with a three- fingered robotic hand, from only a sparse reward and a handful of demonstrations in simulation. Learning from simulation drastically reduces the development cycle of behavior generation, and domain randomization techniques are leveraged to achieve successful zero-shot sim-to- real transfer. Transferred policies are learned directly from raw pixels from multiple cameras and robot proprioception. Our approach outperforms policies learned from demonstrations on the real robot and requires 100 times fewer demonstrations, collected in simulation. More details and videos in sites.google.com/view/demostart.

Many real-world problems require trading off multiple competing objectives. However, these objectives are often in different units and/or scales, which can make it challenging for practitioners to express numerical preferences over objectives in their native units. In this paper we propose a novel algorithm for multi-objective reinforcement learning that enables setting desired preferences for objectives in a scale-invariant way. We propose to learn an action distribution for each objective, and we use supervised learning to fit a parametric policy to a combination of these distributions. We demonstrate the effectiveness of our approach on challenging high-dimensional real and simulated robotics tasks, and show that setting different preferences in our framework allows us to trace out the space of nondominated solutions.

Spatial knowledge plays an essential role in human reasoning, permitting tasks such as locating objects in the world (including oneself), reasoning about everyday actions and describing perceptual information. This is also the case in the field of mobile robotics, where one of the most basic (and essential) tasks is the autonomous determination of the pose of a robot with respect to a map, given its perception of the environment. This is the problem of robot self-localisation (or simply the localisation problem). This paper presents a probabilistic algorithm for robot self-localisation that is based on a topological map constructed from the observation of spatial occlusion. Distinct locations on the map are defined by means of a classical formalism for qualitative spatial reasoning, whose base definitions are closer to the human categorisation of space than traditional, numerical, localisation procedures. The approach herein proposed was systematically evaluated through experiments using a mobile robot equipped with a RGB-D sensor. The results obtained show that the localisation algorithm is successful in locating the robot in qualitatively distinct regions.

This paper presents initial results towards the development of a logic-based probabilistic event recognition system capable of learning and inferring high-level joint actions from simultaneous task execution demonstrations on a search and rescue scenario. We adopt a probabilistic extension of the Event Calculus defined over Markov Logic Networks (MLN-EC). This formalism was applied to learn and infer the actions of human operators teleoperating robots in a real-world robotic search and rescue task. Experimental results in both simulation and real robots show that the probabilistic event logic can recognise the actions taken by the human teleoperators in real world domains containing two collaborating robots, even with uncertain and noisy data.

This paper presents a novel class of algorithms, called Heuristically-Accelerated Multiagent Reinforcement Learning (HAMRL), which allows the use of heuristics to speed up well-known multiagent reinforcement learning (RL) algorithms such as the Minimax-Q. Such HAMRL algorithms are characterized by a heuristic function, which suggests the selection of particular actions over others. This function represents an initial action selection policy, which can be handcrafted, extracted from previous experience in distinct domains, or learnt from observation. To validate the proposal, a thorough theoretical analysis proving the convergence of four algorithms from the HAMRL class (HAMMQ, HAMQ(λ), HAMQS, and HAMS) is presented. In addition, a comprehensive systematical evaluation was conducted in two distinct adversarial domains. The results show that even the most straightforward heuristics can produce virtually optimal action selection policies in much fewer episodes, significantly improving the performance of the HAMRL over vanilla RL algorithms.