Tag Archives: Deep Reinforcement Learning

Using physical models to guide Deep RL in robotics

X. Li, W. Shang and S. Cong, Offline Reinforcement Learning of Robotic Control Using Deep Kinematics and Dynamics, IEEE/ASME Transactions on Mechatronics, vol. 29, no. 4, pp. 2428-2439, Aug. 2024 DOI: 10.1109/TMECH.2023.3336316.

With the rapid development of deep learning, model-free reinforcement learning algorithms have achieved remarkable results in many fields. However, their high sample complexity and the potential for causing damage to environments and robots pose severe challenges for their application in real-world environments. Model-based reinforcement learning algorithms are often used to reduce the sample complexity. One limitation of these algorithms is the inevitable modeling errors. While the black-box model can fit complex state transition models, it ignores the existing knowledge of physics and robotics, especially studies of kinematic and dynamic models of the robotic manipulator. Compared with the black-box model, the physics-inspired deep models do not require specific knowledge of each system to obtain interpretable kinematic and dynamic models. In model-based reinforcement learning, these models can simulate the motion and be combined with classical controllers. This is due to their sharing the same form as traditional models, leading to higher precision tracking results. In this work, we utilize physics-inspired deep models to learn the kinematics and dynamics of a robotic manipulator. We propose a model-based offline reinforcement learning algorithm for controller parameter learning, combined with the traditional computed-torque controller. Experiments on trajectory tracking control of the Baxter manipulator, both in joint and operational space, are conducted in simulation and real environments. Experimental results demonstrate that our algorithm can significantly improve tracking accuracy and exhibits strong generalization and robustness.

Improving reward-sparse situations in RL by adding backward learning

X. Qi, D. Chen, Z. Li and X. Tan, Back-Stepping Experience Replay With Application to Model-Free Reinforcement Learning for a Soft Snake Robot, IEEE Robotics and Automation Letters, vol. 9, no. 9, pp. 7517-7524, Sept. 2024 DOI: 10.1109/LRA.2024.3427550.

In this letter, we propose a novel technique, Back-stepping Experience Replay (BER), that is compatible with arbitrary off-policy reinforcement learning (RL) algorithms. BER aims to enhance learning efficiency in systems with approximate reversibility, reducing the need for complex reward shaping. The method constructs reversed trajectories using back-stepping transitions to reach random or fixed targets. Interpretable as a bi-directional approach, BER addresses inaccuracies in back-stepping transitions through a purification of the replay experience during learning. Given the intricate nature of soft robots and their complex interactions with environments, we present an application of BER in a model-free RL approach for the locomotion and navigation of a soft snake robot, which is capable of serpentine motion enabled by anisotropic friction between the body and ground. In addition, a dynamic simulator is developed to assess the effectiveness and efficiency of the BER algorithm, in which the robot demonstrates successful learning (reaching a 100% success rate) and adeptly reaches random targets, achieving an average speed 48% faster than that of the best baseline approach.

Avoiding the sim-to-real RL transfer problem through learning the parameters of a physical system

Viktor Wiberg, Erik Wallin, Arvid Fälldin, Tobias Semberg, Morgan Rossander, Eddie Wadbro, Martin Servin, Sim-to-real transfer of active suspension control using deep reinforcement learning, Robotics and Autonomous Systems, Volume 179, 2024 DOI: 10.1016/j.robot.2024.104731.

We explore sim-to-real transfer of deep reinforcement learning controllers for a heavy vehicle with active suspensions designed for traversing rough terrain. While related research primarily focuses on lightweight robots with electric motors and fast actuation, this study uses a forestry vehicle with a complex hydraulic driveline and slow actuation. We simulate the vehicle using multibody dynamics and apply system identification to find an appropriate set of simulation parameters. We then train policies in simulation using various techniques to mitigate the sim-to-real gap, including domain randomization, action delays, and a reward penalty to encourage smooth control. In reality, the policies trained with action delays and a penalty for erratic actions perform nearly at the same level as in simulation. In experiments on level ground, the motion trajectories closely overlap when turning to either side, as well as in a route tracking scenario. When faced with a ramp that requires active use of the suspensions, the simulated and real motions are in close alignment. This shows that the actuator model together with system identification yields a sufficiently accurate model of the actuators. We observe that policies trained without the additional action penalty exhibit fast switching or bang–bang control. These present smooth motions and high performance in simulation but transfer poorly to reality. We find that policies make marginal use of the local height map for perception, showing no indications of predictive planning. However, the strong transfer capabilities entail that further development concerning perception and performance can be largely confined to simulation.

Reducing discovered skills in DRL to the essential ones, modelling skills with SMDP Q-learning

Shuai Qing, Fei Zhu, Refine to the essence: Less-redundant skill learning via diversity clustering, Engineering Applications of Artificial Intelligence, Volume 133, Part A, 2024 DOI: 10.1016/j.engappai.2024.107981.

In reinforcement learning, skill is a potentially conditional policy that solves tasks in a hierarchically controlled manner. Progress on skill discovery helps agents learn a set of diverse and useful skills without external supervision to tackle complex tasks with sparse rewards. Although most of the studies have aimed to maximize the diversity of skills discovered, the distinguishability between skills diminishes as the number of skills increases, leading to a subset of similar and redundant skills. To tackle this problem, a method called Refine to the Essence of Skills (RE-Skill) is proposed, which aims at learning skills with less redundancy. RE-Skill integrates the concepts of cluster analysis and policy distillation, clustering similar skills together based on their unique features, learning the most optimal performance within each cluster, and filtering out similar skills that involve excessive and intricate actions, thereby reducing redundancy among skills. By refining clusters of similar skills into less-redundant independent skills, RE-Skill demonstrates superior performance compared to other skill discovery algorithms and shows how these less-redundant skills effectively address downstream tasks, indicating that RE-Skill is able to extend its efficacy to engineering applications in robot control and obstacle training tasks within complex environments.

On the influence of the representations obtained through Deep RL in the learning process

Han Wang, Erfan Miahi, Martha White, Marlos C. Machado, Zaheer Abbas, Raksha Kumaraswamy, Vincent Liu, Adam White, Investigating the properties of neural network representations in reinforcement learning, Artificial Intelligence, Volume 330, 2024 DOI: 10.1016/j.artint.2024.104100.

In this paper we investigate the properties of representations learned by deep reinforcement learning systems. Much of the early work on representations for reinforcement learning focused on designing fixed-basis architectures to achieve properties thought to be desirable, such as orthogonality and sparsity. In contrast, the idea behind deep reinforcement learning methods is that the agent designer should not encode representational properties, but rather that the data stream should determine the properties of the representation—good representations emerge under appropriate training schemes. In this paper we bring these two perspectives together, empirically investigating the properties of representations that support transfer in reinforcement learning. We introduce and measure six representational properties over more than 25,000 agent-task settings. We consider Deep Q-learning agents with different auxiliary losses in a pixel-based navigation environment, with source and transfer tasks corresponding to different goal locations. We develop a method to better understand why some representations work better for transfer, through a systematic approach varying task similarity and measuring and correlating representation properties with transfer performance. We demonstrate the generality of the methodology by investigating representations learned by a Rainbow agent that successfully transfers across Atari 2600 game modes.

Graph NNs in RL for improving sample efficiency

Feng Zhang, Chengbin Xuan, Hak-Keung Lam, An obstacle avoidance-specific reinforcement learning method based on fuzzy attention mechanism and heterogeneous graph neural networks, Engineering Applications of Artificial Intelligence, Volume 130, 2024 DOI: 10.1016/j.engappai.2023.107764.

Deep reinforcement learning (RL) is an advancing learning tool to handle robotics control problems. However, it typically suffers from sample efficiency and effectiveness. The emergence of Graph Neural Networks (GNNs) enables the integration of the RL and graph representation learning techniques. It realises outstanding training performance and transfer capability by forming controlling scenarios into the corresponding graph domain. Nevertheless, the existing approaches strongly depend on the artificial graph formation processes with intensive bias and cannot propagate messages discriminatively on explicit physical dependence, which leads to restricted flexibility, size transfer capability and suboptimal performance. This paper proposes a fuzzy attention mechanism-based heterogeneous graph neural network (FAM-HGNN) framework for resolving the control problem under the RL context. FAM emphasises the significant connections and weakening of the trivial connections in a fully connected graph, which mitigates the potential negative influence caused by the artificial graph formation process. HGNN obtains a higher level of relational inductive bias by conducting graph propagations on a masked graph. Experimental results show that our FAM-HGNN outperforms the multi-layer perceptron-based and the existing GNN-based RL approaches regarding training performance and size transfer capability. We also conducted an ablation study and sensitivity analysis to validate the efficacy of the proposed method further.

Using Deep RL (TRPO) for selecting best interest points in the environment for path planning

Jie Fan, Xudong Zhang, Yuan Zou, Hierarchical path planner for unknown space exploration using reinforcement learning-based intelligent frontier selection, Expert Systems with Applications, Volume 230, 2023 DOI: 10.1016/j.eswa.2023.120630.

Path planning in unknown environments is extremely useful for some specific tasks, such as exploration of outer space planets, search and rescue in disaster areas, home sweeping services, etc. However, existing frontier-based path planners suffer from insufficient exploration, while reinforcement learning (RL)-based ones are confronted with problems in efficient training and effective searching. To overcome the above problems, this paper proposes a novel hierarchical path planner for unknown space exploration using RL-based intelligent frontier selection. Firstly, by decomposing the path planner into three-layered architecture (including the perception layer, planning layer, and control layer) and using edge detection to find potential frontiers to track, the path search space is shrunk from the whole map to a handful of points of interest, which significantly saves the computational resources in both training and execution processes. Secondly, one of the advanced RL algorithms, trust region policy optimization (TRPO), is used as a judge to select the best frontier for the robot to track, which ensures the optimality of the path planner with a shorter path length. The proposed method is validated through simulation and compared with both classic and state-of-the-art methods. Results show that the training process could be greatly accelerated compared with the traditional deep-Q network (DQN). Moreover, the proposed method has 4.2%\u201314.3% improvement in exploration region rate and achieves the highest exploration completeness.

Improving safety in deep RL in the case of autonomous driving

Eduardo Candela, Olivier Doustaly, Leandro Parada, Felix Feng, Yiannis Demiris, Panagiotis Angeloudis, Risk-aware controller for autonomous vehicles using model-based collision prediction and reinforcement learning, Artificial Intelligence, Volume 320, 2023 DOI: 10.1016/j.artint.2023.103923.

Autonomous Vehicles (AVs) have the potential to save millions of lives and increase the efficiency of transportation services. However, the successful deployment of AVs requires tackling multiple challenges related to modeling and certifying safety. State-of-the-art decision-making methods usually rely on end-to-end learning or imitation learning approaches, which still pose significant safety risks. Hence the necessity of risk-aware AVs that can better predict and handle dangerous situations. Furthermore, current approaches tend to lack explainability due to their reliance on end-to-end Deep Learning, where significant causal relationships are not guaranteed to be learned from data. This paper introduces a novel risk-aware framework for training AV agents using a bespoke collision prediction model and Reinforcement Learning (RL). The collision prediction model is based on Gaussian Processes and vehicle dynamics, and is used to generate the RL state vector. Using an explicit risk model increases the post-hoc explainability of the AV agent, which is vital for reaching and certifying the high safety levels required for AVs and other safety-sensitive applications. Experimental results obtained with a simulator and state-of-the-art RL algorithms show that the risk-aware RL framework decreases average collision rates by 15%, makes AVs more robust to sudden harsh braking situations, and achieves better performance in both safety and speed when compared to a standard rule-based method (the Intelligent Driver Model). Moreover, the proposed collision prediction model outperforms other models in the literature.

See also: https://doi.org/10.1016/j.artint.2023.103922
And also: https://doi.org/10.1177/02783649231169492

Doing a more intelligent exploration in RL based on measuring uncertainty through prediction

Xiaoshu Zhou, Fei Zhu, Peiyao Zhao, Within the scope of prediction: Shaping intrinsic rewards via evaluating uncertainty, Expert Systems with Applications, Volume 206, 2022 DOI: 10.1016/j.eswa.2022.117775.

The agent of reinforcement learning based approaches needs to explore to learn more about the environment to seek optimal policy. However, simply increasing the frequency of stochastic exploration sometimes fails to work or even causes the agent to fall into traps. To solve the problem, it is essential to improve the quality of exploration. An approach, referred to as the scope of prediction based on uncertainty exploration (SPE), is proposed, taking advantage of the uncertainty mechanism and considering the stochasticity of prospecting. As by uncertainty mechanism, the unexpected states make more curiosity, the model derives higher uncertainty by projecting future scenarios to compare with the actual future to explore the world. The SPE method utilizes a prediction network to predict subsequent observations and calculates the mean squared difference value of the real observations and the following observations to measure uncertainty, encouraging the agent to explore unknown regions more effectively. Moreover, to reduce the noise interference caused by uncertainty, a reward-penalty model is developed to discriminate the noise by current observations and action prediction for future rewards to improve the interference ability against noise so that the agent can escape from the noisy region. Experiment results showed that deep reinforcement learning approaches equipped with SPE demonstrated significant improvements in simulated environments.

Improving the quality of memory replay in RL through an evolutionary algorithm biologically inspired

M. Ramicic and A. Bonarini, Augmented Memory Replay in Reinforcement Learning With Continuous Control, IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 2, pp. 485-496, June 2022 DOI: 10.1109/TCDS.2021.3050723.

Online reinforcement learning agents are currently able to process an increasing amount of data by converting it into a higher order value functions. This expansion of the information collected from the environment increases the agent\u2019s state space enabling it to scale up to more complex problems but also increases the risk of forgetting by learning on redundant or conflicting data. To improve the approximation of a large amount of data, a random mini-batch of the past experiences that are stored in the replay memory buffer is often replayed at each learning step. The proposed work takes inspiration from a biological mechanism which acts as a protective layer of higher cognitive functions found in mammalian brain: active memory consolidation mitigates the effect of forgetting previous memories by dynamically processing the new ones. Similar dynamics are implemented by the proposed augmented memory replay or AMR algorithm. The architecture of AMR , based on a simple artificial neural network is able to provide an augmentation policy which modifies each of the agents experiences by augmenting their relevance prior to storing them in the replay memory. The function approximator of AMR is evolved using genetic algorithm in order to obtain the specific augmentation policy function that yields the best performance of a learning agent in a specific environment given by its received cumulative reward. Experimental results show that an evolved AMR augmentation function capable of increasing the significance of the specific memories is able to further increase the stability and convergence speed of the learning algorithms dealing with the complexity of continuous action domains.