Takamitsu Matsubara

Yuki Kadokawa, Lingwei Zhu, Yoshihisa Tsurumine et al.

Deep reinforcement learning with domain randomization learns a control policy in various simulations with randomized physical and sensor model parameters to become transferable to the real world in a zero-shot setting. However, a huge number of samples are often required to learn an effective policy when the range of randomized parameters is extensive due to the instability of policy updates. To alleviate this problem, we propose a sample-efficient method named cyclic policy distillation (CPD). CPD divides the range of randomized parameters into several small sub-domains and assigns a local policy to each one. Then local policies are learned while cyclically transitioning to sub-domains. CPD accelerates learning through knowledge transfer based on expected performance improvements. Finally, all of the learned local policies are distilled into a global policy for sim-to-real transfers. CPD's effectiveness and sample efficiency are demonstrated through simulations with four tasks (Pendulum from OpenAIGym and Pusher, Swimmer, and HalfCheetah from Mujoco), and a real-robot, ball-dispersal task. We published code and videos from our experiments at https://github.com/yuki-kadokawa/cyclic-policy-distillation.

Tomoya Yamanokuchi, Yuhwan Kwon, Yoshihisa Tsurumine et al.

Many works have recently explored Sim-to-real transferable visual model predictive control (MPC). However, such works are limited to one-shot transfer, where real-world data must be collected once to perform the sim-to-real transfer, which remains a significant human effort in transferring the models learned in simulations to new domains in the real world. To alleviate this problem, we first propose a novel model-learning framework called Kalman Randomized-to-Canonical Model (KRC-model). This framework is capable of extracting task-relevant intrinsic features and their dynamics from randomized images. We then propose Kalman Randomized-to-Canonical Model Predictive Control (KRC-MPC) as a zero-shot sim-to-real transferable visual MPC using KRC-model. The effectiveness of our method is evaluated through a valve rotation task by a robot hand in both simulation and the real world, and a block mating task in simulation. The experimental results show that KRC-MPC can be applied to various real domains and tasks in a zero-shot manner.

Hirotaka Tahara, Hikaru Sasaki, Hanbit Oh et al.

Partial Automation (PA) with intelligent support systems has been introduced in industrial machinery and advanced automobiles to reduce the burden of long hours of human operation. Under PA, operators perform manual operations (providing actions) and operations that switch to automatic/manual mode (mode-switching). Since PA reduces the total duration of manual operation, these two action and mode-switching operations can be replicated by imitation learning with high sample efficiency. To this end, this paper proposes Disturbance Injection under Partial Automation (DIPA) as a novel imitation learning framework. In DIPA, mode and actions (in the manual mode) are assumed to be observables in each state and are used to learn both action and mode-switching policies. The above learning is robustified by injecting disturbances into the operator's actions to optimize the disturbance's level for minimizing the covariate shift under PA. We experimentally validated the effectiveness of our method for long-horizon tasks in two simulations and a real robot environment and confirmed that our method outperformed the previous methods and reduced the demonstration burden.

Ngoc Huy Nguyen, Kazuki Shibata, Takamitsu Matsubara

In this study, we address the problem of in-flight object catching using a quadruped robot with a basket. Our objective is to accurately predict the impact point, defined as the object's landing position. This task poses two key challenges: the absence of public datasets capturing diverse objects under unsteady aerodynamics, which are essential for training reliable predictors; and the difficulty of accurate early-stage impact point prediction when trajectories appear similar across objects. To overcome these issues, we construct a real-world dataset of 8,000 trajectories from 20 objects, providing a foundation for advancing in-flight object catching under complex aerodynamics. We then propose the Object-Adaptive Impact Point Predictor (OIPP), consisting of two modules: (i) an Object-Adaptive Encoder (OAE) that extracts object-dependent representations from motion histories, and (ii) an Impact Point Predictor (IPP) that estimates the impact point from these representations. Two IPP variants are implemented: a Neural Acceleration Estimator (NAE)-based method that predicts trajectories and derives the impact point, and a Direct Point Estimator (DPE)-based method that directly outputs it. Experimental results show that our dataset is more diverse and complex than existing datasets, and that our method outperforms baselines on both 15 seen and 5 unseen objects. Furthermore, we show that improved early-stage prediction enhances catching success in simulation and demonstrate the effectiveness of our approach through real-robot experiments. The demonstration is available at https://sites.google.com/view/robot-catching-2025.

Lingwei Zhu, Zheng Chen, Eiji Uchibe et al.

Maximum Tsallis entropy (MTE) framework in reinforcement learning has gained popularity recently by virtue of its flexible modeling choices including the widely used Shannon entropy and sparse entropy. However, non-Shannon entropies suffer from approximation error and subsequent underperformance either due to its sensitivity or the lack of closed-form policy expression. To improve the tradeoff between flexibility and empirical performance, we propose to strengthen their error-robustness by enforcing implicit Kullback-Leibler (KL) regularization in MTE motivated by Munchausen DQN (MDQN). We do so by drawing connection between MDQN and advantage learning, by which MDQN is shown to fail on generalizing to the MTE framework. The proposed method Tsallis Advantage Learning (TAL) is verified on extensive experiments to not only significantly improve upon Tsallis-DQN for various non-closed-form Tsallis entropies, but also exhibits comparable performance to state-of-the-art maximum Shannon entropy algorithms.

Lingwei Zhu, Zheng Chen, Eiji Uchibe et al.

The recently successful Munchausen Reinforcement Learning (M-RL) features implicit Kullback-Leibler (KL) regularization by augmenting the reward function with logarithm of the current stochastic policy. Though significant improvement has been shown with the Boltzmann softmax policy, when the Tsallis sparsemax policy is considered, the augmentation leads to a flat learning curve for almost every problem considered. We show that it is due to the mismatch between the conventional logarithm and the non-logarithmic (generalized) nature of Tsallis entropy. Drawing inspiration from the Tsallis statistics literature, we propose to correct the mismatch of M-RL with the help of $q$-logarithm/exponential functions. The proposed formulation leads to implicit Tsallis KL regularization under the maximum Tsallis entropy framework. We show such formulation of M-RL again achieves superior performance on benchmark problems and sheds light on more general M-RL with various entropic indices $q$.

Yuki Kadokawa, Hirotaka Tahara, Takamitsu Matsubara

In earthwork and construction, excavators often encounter large rocks mixed with various soil conditions, requiring skilled operators. This paper presents a framework for achieving autonomous excavation using reinforcement learning (RL) through a rock excavation simulator. In the simulation, resolution can be defined by the particle size/number in the whole soil space. Fine-resolution simulations closely mimic real-world behavior but demand significant calculation time and challenging sample collection, while coarse-resolution simulations enable faster sample collection but deviate from real-world behavior. To combine the advantages of both resolutions, we explore using policies developed in coarse-resolution simulations for pre-training in fine-resolution simulations. To this end, we propose a novel policy learning framework called Progressive-Resolution Policy Distillation (PRPD), which progressively transfers policies through some middle-resolution simulations with conservative policy transfer to avoid domain gaps that could lead to policy transfer failure. Validation in a rock excavation simulator and nine real-world rock environments demonstrated that PRPD reduced sampling time to less than 1/7 while maintaining task success rates comparable to those achieved through policy learning in a fine-resolution simulation.

Gahee Kim, Takamitsu Matsubara

Black-box simulators are widely used in robotics, but optimizing their parameters remains challenging due to inaccessible likelihoods. Simulation-Based Inference (SBI) tackles this issue using simulation-driven approaches, estimating the posterior from offline real observations and forward simulations. However, in black-box scenarios, preparing observations that contain sufficient information for parameter estimation is difficult due to the unknown relationship between parameters and observations. In this work, we present Active Simulation-Based Inference (ASBI), a parameter estimation framework that uses robots to actively collect real-world online data to achieve accurate black-box simulator tuning. Our framework optimizes robot actions to collect informative observations by maximizing information gain, which is defined as the expected reduction in Shannon entropy between the posterior and the prior. While calculating information gain requires the likelihood, which is inaccessible in black-box simulators, our method solves this problem by leveraging Neural Posterior Estimation (NPE), which leverages a neural network to learn the posterior estimator. Three simulation experiments quantitatively verify that our method achieves accurate parameter estimation, with posteriors sharply concentrated around the true parameters. Moreover, we show a practical application using a real robot to estimate the simulation parameters of cubic particles corresponding to two real objects, beads and gravel, with a bucket pouring action.

Po-Yen Wu, Cheng-Yu Kuo, Yuki Kadokawa et al.

In inaccessible environments with uncertain task demands, robots often rely on general-purpose tools that lack predefined usage strategies. These tools are not tailored for particular operations, making their longevity highly sensitive to how they are used. This creates a fundamental challenge: how can a robot learn a tool-use policy that both completes the task and prolongs the tool's lifespan? In this work, we address this challenge by introducing a reinforcement learning (RL) framework that incorporates tool lifespan as a factor during policy optimization. Our framework leverages Finite Element Analysis (FEA) and Miner's Rule to estimate Remaining Useful Life (RUL) based on accumulated stress, and integrates the RUL into the RL reward to guide policy learning toward lifespan-guided behavior. To handle the fact that RUL can only be estimated after task execution, we introduce an Adaptive Reward Normalization (ARN) mechanism that dynamically adjusts reward scaling based on estimated RULs, ensuring stable learning signals. We validate our method across simulated and real-world tool use tasks, including Object-Moving and Door-Opening with multiple general-purpose tools. The learned policies consistently prolong tool lifespan (up to 8.01x in simulation) and transfer effectively to real-world settings, demonstrating the practical value of learning lifespan-guided tool use strategies.

Hirotaka Tahara, Takamitsu Matsubara

Applying imitation learning (IL) is challenging to nonprehensile manipulation tasks of invisible objects with partial observations, such as excavating buried rocks. The demonstrator must make such complex action decisions as exploring to find the object and task-oriented actions to complete the task while estimating its hidden state, perhaps causing inconsistent action demonstration and high cognitive load problems. For these problems, work in human cognitive science suggests that promoting the use of pre-designed, simple exploration rules for the demonstrator may alleviate the problems of action inconsistency and high cognitive load. Therefore, when performing imitation learning from demonstrations using such exploration rules, it is important to accurately imitate not only the demonstrator's task-oriented behavior but also his/her mode-switching behavior (exploratory or task-oriented behavior) under partial observation. Based on the above considerations, this paper proposes a novel imitation learning framework called Belief Exploration-Action Cloning (BEAC), which has a switching policy structure between a pre-designed exploration policy and a task-oriented action policy trained on the estimated belief states based on past history. In simulation and real robot experiments, we confirmed that our proposed method achieved the best task performance, higher mode and action prediction accuracies, while reducing the cognitive load in the demonstration indicated by a user study.

Kei Takahashi, Hikaru Sasaki, Takamitsu Matsubara

Imitation learning through a demonstration interface is expected to learn policies for robot automation from intuitive human demonstrations. However, due to the differences in human and robot movement characteristics, a human expert might unintentionally demonstrate an action that the robot cannot execute. We propose feasibility-aware behavior cloning from observation (FABCO). In the FABCO framework, the feasibility of each demonstration is assessed using the robot's pre-trained forward and inverse dynamics models. This feasibility information is provided as visual feedback to the demonstrators, encouraging them to refine their demonstrations. During policy learning, estimated feasibility serves as a weight for the demonstration data, improving both the data efficiency and the robustness of the learned policy. We experimentally validated FABCO's effectiveness by applying it to a pipette insertion task involving a pipette and a vial. Four participants assessed the impact of the feasibility feedback and the weighted policy learning in FABCO. Additionally, we used the NASA Task Load Index (NASA-TLX) to evaluate the workload induced by demonstrations with visual feedback.

Shu-yuan Wang, Hikaru Sasaki, Takamitsu Matsubara

Learning control policies for real-world robotic tasks often involve challenges such as multimodality, local discontinuities, and the need for computational efficiency. These challenges arise from the complexity of robotic environments, where multiple solutions may coexist. To address these issues, we propose Composite Gaussian Processes Flows (CGP-Flows), a novel semi-parametric model for robotic policy. CGP-Flows integrate Overlapping Mixtures of Gaussian Processes (OMGPs) with the Continuous Normalizing Flows (CNFs), enabling them to model complex policies addressing multimodality and local discontinuities. This hybrid approach retains the computational efficiency of OMGPs while incorporating the flexibility of CNFs. Experiments conducted in both simulated and real-world robotic tasks demonstrate that CGP-flows significantly improve performance in modeling control policies. In a simulation task, we confirmed that CGP-Flows had a higher success rate compared to the baseline method, and the success rate of GCP-Flow was significantly different from the success rate of other baselines in chi-square tests.

Keiichiro Takahashi, Taisuke Kobayashi, Tomoya Yamanokuchi et al.

This paper investigates a novel nonlinear update rule based on temporal difference (TD) errors in reinforcement learning (RL). The update rule in the standard RL states that the TD error is linearly proportional to the degree of updates, treating all rewards equally without no bias. On the other hand, the recent biological studies revealed that there are nonlinearities in the TD error and the degree of updates, biasing policies optimistic or pessimistic. Such biases in learning due to nonlinearities are expected to be useful and intentionally leftover features in biological learning. Therefore, this research explores a theoretical framework that can leverage the nonlinearity between the degree of the update and TD errors. To this end, we focus on a control as inference framework, since it is known as a generalized formulation encompassing various RL and optimal control methods. In particular, we investigate the uncomputable nonlinear term needed to be approximately excluded in the derivation of the standard RL from control as inference. By analyzing it, Weber-Fechner law (WFL) is found, namely, perception (a.k.a. the degree of updates) in response to stimulus change (a.k.a. TD error) is attenuated by increase in the stimulus intensity (a.k.a. the value function). To numerically reveal the utilities of WFL on RL, we then propose a practical implementation using a reward-punishment framework and modifying the definition of optimality. Analysis of this implementation reveals that two utilities can be expected i) to increase rewards to a certain level early, and ii) to sufficiently suppress punishment. We finally investigate and discuss the expected utilities through simulations and robot experiments. As a result, the proposed RL algorithm with WFL shows the expected utilities that accelerate the reward-maximizing startup and continue to suppress punishments during learning.

Takuya Kiyokawa, Eiki Nagata, Yoshihisa Tsurumine et al.

Mobile grasping enhances manipulation efficiency by utilizing robots' mobility. This study aims to enable a commercial off-the-shelf robot for mobile grasping, requiring precise timing and pose adjustments. Self-supervised learning can develop a generalizable policy to adjust the robot's velocity and determine grasp position and orientation based on the target object's shape and pose. Due to mobile grasping's complexity, action primitivization and step-by-step learning are crucial to avoid data sparsity in learning from trial and error. This study simplifies mobile grasping into two grasp action primitives and a moving action primitive, which can be operated with limited degrees of freedom for the manipulator. This study introduces three fully convolutional neural network (FCN) models to predict static grasp primitive, dynamic grasp primitive, and residual moving velocity error from visual inputs. A two-stage grasp learning approach facilitates seamless FCN model learning. The ablation study demonstrated that the proposed method achieved the highest grasping accuracy and pick-and-place efficiency. Furthermore, randomizing object shapes and environments in the simulation effectively achieved generalizable mobile grasping.

Wendyam Eric Lionel Ilboudo, Taisuke Kobayashi, Takamitsu Matsubara

With the increasing practicality of deep learning applications, practitioners are inevitably faced with datasets corrupted by noise from various sources such as measurement errors, mislabeling, and estimated surrogate inputs/outputs that can adversely impact the optimization results. It is a common practice to improve the optimization algorithm's robustness to noise, since this algorithm is ultimately in charge of updating the network parameters. Previous studies revealed that the first-order moment used in Adam-like stochastic gradient descent optimizers can be modified based on the Student's t-distribution. While this modification led to noise-resistant updates, the other associated statistics remained unchanged, resulting in inconsistencies in the assumed models. In this paper, we propose AdaTerm, a novel approach that incorporates the Student's t-distribution to derive not only the first-order moment but also all the associated statistics. This provides a unified treatment of the optimization process, offering a comprehensive framework under the statistical model of the t-distribution for the first time. The proposed approach offers several advantages over previously proposed approaches, including reduced hyperparameters and improved robustness and adaptability. This noise-adaptive behavior contributes to AdaTerm's exceptional learning performance, as demonstrated through various optimization problems with different and/or unknown noise ratios. Furthermore, we introduce a new technique for deriving a theoretical regret bound without relying on AMSGrad, providing a valuable contribution to the field

Yuki Kadokawa, Yoshihisa Tsurumine, Takamitsu Matsubara

This paper explores a Deep Reinforcement Learning (DRL) approach for designing image-based control for edge robots to be implemented on Field Programmable Gate Arrays (FPGAs). Although FPGAs are more power-efficient than CPUs and GPUs, a typical DRL method cannot be applied since they are composed of many Logic Blocks (LBs) for high-speed logical operations but low-speed real-number operations. To cope with this problem, we propose a novel DRL algorithm called Binarized P-Network (BPN), which learns image-input control policies using Binarized Convolutional Neural Networks (BCNNs). To alleviate the instability of reinforcement learning caused by a BCNN with low function approximation accuracy, our BPN adopts a robust value update scheme called Conservative Value Iteration, which is tolerant of function approximation errors. We confirmed the BPN's effectiveness through applications to a visual tracking task in simulation and real-robot experiments with FPGA.

Toshinori Kitamura, Lingwei Zhu, Takamitsu Matsubara

The recent boom in the literature on entropy-regularized reinforcement learning (RL) approaches reveals that Kullback-Leibler (KL) regularization brings advantages to RL algorithms by canceling out errors under mild assumptions. However, existing analyses focus on fixed regularization with a constant weighting coefficient and do not consider cases where the coefficient is allowed to change dynamically. In this paper, we study the dynamic coefficient scheme and present the first asymptotic error bound. Based on the dynamic coefficient error bound, we propose an effective scheme to tune the coefficient according to the magnitude of error in favor of more robust learning. Complementing this development, we propose a novel algorithm, Geometric Value Iteration (GVI), that features a dynamic error-aware KL coefficient design with the aim of mitigating the impact of errors on performance. Our experiments demonstrate that GVI can effectively exploit the trade-off between learning speed and robustness over uniform averaging of a constant KL coefficient. The combination of GVI and deep networks shows stable learning behavior even in the absence of a target network, where algorithms with a constant KL coefficient would greatly oscillate or even fail to converge.

Lingwei Zhu, Toshinori Kitamura, Takamitsu Matsubara

In this paper, we propose cautious policy programming (CPP), a novel value-based reinforcement learning (RL) algorithm that can ensure monotonic policy improvement during learning. Based on the nature of entropy-regularized RL, we derive a new entropy regularization-aware lower bound of policy improvement that only requires estimating the expected policy advantage function. CPP leverages this lower bound as a criterion for adjusting the degree of a policy update for alleviating policy oscillation. Different from similar algorithms that are mostly theory-oriented, we also propose a novel interpolation scheme that makes CPP better scale in high dimensional control problems. We demonstrate that the proposed algorithm can trade o? performance and stability in both didactic classic control problems and challenging high-dimensional Atari games.

Lingwei Zhu, Toshinori Kitamura, Takamitsu Matsubara

The oscillating performance of off-policy learning and persisting errors in the actor-critic (AC) setting call for algorithms that can conservatively learn to suit the stability-critical applications better. In this paper, we propose a novel off-policy AC algorithm cautious actor-critic (CAC). The name cautious comes from the doubly conservative nature that we exploit the classic policy interpolation from conservative policy iteration for the actor and the entropy-regularization of conservative value iteration for the critic. Our key observation is the entropy-regularized critic facilitates and simplifies the unwieldy interpolated actor update while still ensuring robust policy improvement. We compare CAC to state-of-the-art AC methods on a set of challenging continuous control problems and demonstrate that CAC achieves comparable performance while significantly stabilizes learning.

Hikaru Sasaki, Takamitsu Matsubara

Policy search reinforcement learning has been drawing much attention as a method of learning a robot control policy. In particular, policy search using such non-parametric policies as Gaussian process regression can learn optimal actions with high-dimensional and redundant sensors as input. However, previous methods implicitly assume that the optimal action becomes unique for each state. This assumption can severely limit such practical applications as robot manipulations since designing a reward function that appears in only one optimal action for complex tasks is difficult. The previous methods might have caused critical performance deterioration because the typical non-parametric policies cannot capture the optimal actions due to their unimodality. We propose novel approaches in non-parametric policy searches with multiple optimal actions and offer two different algorithms commonly based on a sparse Gaussian process prior and variational Bayesian inference. The following are the key ideas: 1) multimodality for capturing multiple optimal actions and 2) mode-seeking for capturing one optimal action by ignoring the others. First, we propose a multimodal sparse Gaussian process policy search that uses multiple overlapped GPs as a prior. Second, we propose a mode-seeking sparse Gaussian process policy search that uses the student-t distribution for a likelihood function. The effectiveness of those algorithms is demonstrated through applications to object manipulation tasks with multiple optimal actions in simulations.

Kazuki Shibata, Tatsuya Miyano, Tomohiko Jimbo et al.

We propose a means of omni-directional contact detection using accelerometers instead of tactile sensors for object shape estimation using touch. Unlike tactile sensors, our contact-based detection method tends to induce a degree of uncertainty with false-positive contact data because the sensors may react not only to actual contact but also to the unstable behavior of the robot. Therefore, it is crucial to consider a robust shape estimation method capable of handling such false-positive contact data. To realize this, we introduce the concept of heteroscedasticity into the contact data and propose a robust shape estimation algorithm based on Gaussian process implicit surfaces (GPIS). We confirmed that our algorithm not only reduces shape estimation errors caused by false-positive contact data but also distinguishes false-positive contact data more clearly than the GPIS through simulations and actual experiments using a quadcopter.

Naoto Komeno, Takamitsu Matsubara

Tactile perception using vibration sensation helps robots recognize their environment's physical properties and perform complex tasks. A sliding motion is applied to target objects to generate tactile vibration data. However, situations exist where such a sliding motion is infeasible due to geometrical constraints in the environment or an object's fragility which cannot resist friction forces. This paper explores a novel approach to achieve vibration-based tactile perception without a sliding motion. To this end, our key idea is injecting a mechanical vibration into a soft tactile sensor system and measuring the propagated vibration inside it by a sensor. Soft tactile sensors are deformed by the contact state, and the touched objects' shape or texture should change the characteristics of the vibration propagation. Therefore, the propagated-vibration data are expected to contain useful information for recognizing touched environments. We developed a prototype system for a proof-of-concept: a mechanical vibration is applied to a biomimetic (soft and vibration-based) tactile sensor from a small, mounted piezoelectric actuator. As a verification experiment, we performed two classification tasks for sandpaper's grit size and a slit's gap widths using our approach and compared their accuracies with that of using sliding motions. Our approach resulted in 70% accuracy for the grit size classification and 99% accuracy for the gap width classification. These results are comparable to or better than the comparison methods with a sliding motion.

Kazuki Shibata, Tomohiko Jimbo, Takamitsu Matsubara

In this paper, we explore a multi-agent reinforcement learning approach to address the design problem of communication and control strategies for multi-agent cooperative transport. Typical end-to-end deep neural network policies may be insufficient for covering communication and control; these methods cannot decide the timing of communication and can only work with fixed-rate communications. Therefore, our framework exploits event-triggered architecture, namely, a feedback controller that computes the communication input and a triggering mechanism that determines when the input has to be updated again. Such event-triggered control policies are efficiently optimized using a multi-agent deep deterministic policy gradient. We confirmed that our approach could balance the transport performance and communication savings through numerical simulations.

Hanbit Oh, Hikaru Sasaki, Brendan Michael et al.

Scenarios requiring humans to choose from multiple seemingly optimal actions are commonplace, however standard imitation learning often fails to capture this behavior. Instead, an over-reliance on replicating expert actions induces inflexible and unstable policies, leading to poor generalizability in an application. To address the problem, this paper presents the first imitation learning framework that incorporates Bayesian variational inference for learning flexible non-parametric multi-action policies, while simultaneously robustifying the policies against sources of error, by introducing and optimizing disturbances to create a richer demonstration dataset. This combinatorial approach forces the policy to adapt to challenging situations, enabling stable multi-action policies to be learned efficiently. The effectiveness of our proposed method is evaluated through simulations and real-robot experiments for a table-sweep task using the UR3 6-DOF robotic arm. Results show that, through improved flexibility and robustness, the learning performance and control safety are better than comparison methods.

Cheng-Yu Kuo, Andreas Schaarschmidt, Yunduan Cui et al.

This letter presents contact-safe Model-based Reinforcement Learning (MBRL) for robot applications that achieves contact-safe behaviors in the learning process. In typical MBRL, we cannot expect the data-driven model to generate accurate and reliable policies to the intended robotic tasks during the learning process due to sample scarcity. Operating these unreliable policies in a contact-rich environment could cause damage to the robot and its surroundings. To alleviate the risk of causing damage through unexpected intensive physical contacts, we present the contact-safe MBRL that associates the probabilistic Model Predictive Control's (pMPC) control limits with the model uncertainty so that the allowed acceleration of controlled behavior is adjusted according to learning progress. Control planning with such uncertainty-aware control limits is formulated as a deterministic MPC problem using a computation-efficient approximated GP dynamics and an approximated inference technique. Our approach's effectiveness is evaluated through bowl mixing tasks with simulated and real robots, scooping tasks with a real robot as examples of contact-rich manipulation skills. (video: https://youtu.be/sdhHP3NhYi0)

Lingwei Zhu, Takamitsu Matsubara

This paper aims to establish an entropy-regularized value-based reinforcement learning method that can ensure the monotonic improvement of policies at each policy update. Unlike previously proposed lower-bounds on policy improvement in general infinite-horizon MDPs, we derive an entropy-regularization aware lower bound. Since our bound only requires the expected policy advantage function to be estimated, it is scalable to large-scale (continuous) state-space problems. We propose a novel reinforcement learning algorithm that exploits this lower-bound as a criterion for adjusting the degree of a policy update for alleviating policy oscillation. We demonstrate the effectiveness of our approach in both discrete-state maze and continuous-state inverted pendulum tasks using a linear function approximator for value estimation.

Cristian Camilo Beltran-Hernandez, Damien Petit, Ixchel G. Ramirez-Alpizar et al.

Reinforcement Learning (RL) methods have been proven successful in solving manipulation tasks autonomously. However, RL is still not widely adopted on real robotic systems because working with real hardware entails additional challenges, especially when using rigid position-controlled manipulators. These challenges include the need for a robust controller to avoid undesired behavior, that risk damaging the robot and its environment, and constant supervision from a human operator. The main contributions of this work are, first, we proposed a learning-based force control framework combining RL techniques with traditional force control. Within said control scheme, we implemented two different conventional approaches to achieve force control with position-controlled robots; one is a modified parallel position/force control, and the other is an admittance control. Secondly, we empirically study both control schemes when used as the action space of the RL agent. Thirdly, we developed a fail-safe mechanism for safely training an RL agent on manipulation tasks using a real rigid robot manipulator. The proposed methods are validated on simulation and a real robot, an UR3 e-series robotic arm.

Takamitsu Matsubara, Vicenç Gómez, Hilbert J. Kappen

Kullback Leibler (KL) control problems allow for efficient computation of optimal control by solving a principal eigenvector problem. However, direct applicability of such framework to continuous state-action systems is limited. In this paper, we propose to embed a KL control problem in a probabilistic graphical model where observed variables correspond to the continuous (possibly high-dimensional) state of the system and latent variables correspond to a discrete (low-dimensional) representation of the state amenable for KL control computation. We present two examples of this approach. The first one uses standard hidden Markov models (HMMs) and computes exact optimal control, but is only applicable to low-dimensional systems. The second one uses factorial HMMs, it is scalable to higher dimensional problems, but control computation is approximate. We illustrate both examples in several robot motor control tasks.