1,720,996 research outputs found
A case for robust AI in robotics
No full textResearchers envision a world wherein robots are free to interact with the external environment, thereby including human beings, other living creatures, robots and a variety of inanimate objects. It is always tacitly assumed that interactions will be smooth, i.e., they will fulfill several desirable properties ranging from safety to appropriateness. We posit that a reasonable mathematical model to frame such vision is that of Markov decision processes, and that ensuring smooth interactions amounts to endow robots with control policies that are provably compliant with side conditions expressed in probabilistic temporal logic
Evaluating probabilistic model checking tools for verification of robot control policies
No full textResearch literature on Probabilistic Model Checking (PMC) encompasses a well-established set of algorithmic techniques whereby probabilistic models can be analyzed. In the last decade, owing to the increasing availability of effective tools, PMC has found applications in many domains, including computer networks, computational biology and robotics. In this paper, we evaluate PMC tools -namely comics, mrmc and prism -to investigate safe reinforcement learning in robots, i.e., to establish safety of policies learned considering feedback signals received upon acting in partially unknown environments. Introduced in previous contributions of ours, this application is a challenging domain wherein PMC tools act as back-engines of an automated methodology aimed to verify and repair control policies. We present an evaluation of the current state-of-the-art PMC tools to assess their potential on various case studies, including both real and simulated robots accomplishing navigation, manipulation and reaching tasks
Testing a Learn-Verify-Repair Approach for Safe Human-Robot Interaction
No full textEnsuring safe behaviors, i.e., minimizing the probability that a control strategy yields undesirable effects, becomes crucial when robots interact with humans in semi-structured environments through adaptive control strategies. In previous papers, we contributed to propose an approach that (i) computes control policies through reinforcement learning, (ii) verifies them against safety requirements with probabilistic model checking, and (iii) repairs them with greedy local methods until requirements are met. Such learn-verify-repair work-flow was shown effective in some — relatively simple and confined — test cases. In this paper, we frame human-robot interaction in light of such previous contributions, and we test the effectiveness of the learn-verify-repair approach in a more realistic factory-to-home deployment scenario. The purpose of our test is to assess whether we can verify that interaction patterns are carried out with negligible human-to-robot collision probability and whether, in the presence of user tuning, strategies which determine offending behaviors can be effectively repaired
Verification and repair of control policies for safe reinforcement learning
No full textReinforcement Learning is a well-known AI paradigm whereby control policies of autonomous agents can be synthesized in an incremental fashion with little or no knowledge about the properties of the environment. We are concerned with safety of agents whose policies are learned by reinforcement, i.e., we wish to bound the risk that, once learning is over, an agent damages either the environment or itself. We propose a general-purpose automated methodology to verify, i.e., establish risk bounds, and repair policies, i.e., fix policies to comply with stated risk bounds. Our approach is based on probabilistic model checking algorithms and tools, which provide theoretical and practical means to verify risk bounds and repair policies. Considering a taxonomy of potential repair approaches tested on an artificially-generated parametric domain, we show that our methodology is also more effective than comparable ones
Is verification a requisite for safe adaptive robots?
No full textThis paper argues in favour of using formal methods to ensure safety of deployed stochastic policies learned by robots in unstructured environments. It has been demonstrated that multi-objective learning alone is not sufficient to ensure globally safe behaviours in such robots, whereas learning-specific methods yield deterministic policies which are less flexible or effective in practice. Under certain restrictions on state-space, modelling safety using probabilistic computational tree logic and ensuring such safety via automated repair can overcome these shortcomings. Promising results are obtained on a realistic setup and pros and cons of such method are discussed
Ensuring safety of policies learned by reinforcement: Reaching objects in the presence of obstacles with the iCub
No full textSafe learning with real-time constraints: a case study
No full textAim of this work is to study the problem of ensuring safety and effectiveness of a multi-agent robot control system with real-time constraints in the case of learning components usage. Our case study focuses on a robot playing the air hockey game against a human opponent, where the robot has to learn how to minimize opponent’s goals. This case study is paradigmatic since the robot must act in real-time, but, at the same time, it must learn and guarantee that the control system is safe throughout the process. We propose a solution using automatatheoretic formalisms and associated verification tools, showing experimentally
that our approach can yield safety without heavily compromising effectiveness
Engineering approaches and methods to verify software in autonomous systems
No full textWe present three computer-augmented software engineering approaches to ensure dependability at different levels of control architectures in autonomous robots. For each approach, we outline the methodological framework, our current achievements, and open issues. Albeit our results are still preliminary, we believe that furthering research along these lines can provide cost-effective techniques to make autonomous robots safe and thus fit for commercial purposes
A Greedy Approach for the Efficient Repair of Stochastic Models
No full textFor discrete-time probabilistic models there are efficient methods to check whether they satisfy certain properties. If a property is refuted, available techniques can be used to explain the failure in form of a counterexample. However, there are no scalable approaches to repair a model, i.e., to modify it with respect to certain side conditions such that the property is satisfied. In this paper we propose such a method, which avoids expensive computations and is therefore applicable to large models. A prototype implementation is used to demonstrate the applicability and scalability of our technique
- …
