Taisuke Kobayashi

Taisuke Kobayashi

In reinforcement learning (RL), temporal difference (TD) error is known to be related to the firing rate of dopamine neurons. It has been observed that each dopamine neuron does not behave uniformly, but each responds to the TD error in an optimistic or pessimistic manner, interpreted as a kind of distributional RL. To explain such a biological data, a heuristic model has also been designed with learning rates asymmetric for the positive and negative TD errors. However, this heuristic model is not theoretically-grounded and unknown whether it can work as a RL algorithm. This paper therefore introduces a novel theoretically-grounded model with optimism and pessimism, which is derived from control as inference. In combination with ensemble learning, a distributional value function as a critic is estimated from regularly introduced optimism and pessimism. Based on its central value, a policy in an actor is improved. This proposed algorithm, so-called DROP (distributional and regular optimism and pessimism), is compared on dynamic tasks. Although the heuristic model showed poor learning performance, DROP showed excellent one in all tasks with high generality. In other words, it was suggested that DROP is a new model that can elicit the potential contributions of optimism and pessimism.

Taisuke Kobayashi

Deep reinforcement learning (DRL) is one of the promising approaches for introducing robots into complicated environments. The recent remarkable progress of DRL stands on regularization of policy, which allows the policy to improve stably and efficiently. A popular method, so-called proximal policy optimization (PPO), and its variants constrain density ratio of the latest and baseline policies when the density ratio exceeds a given threshold. This threshold can be designed relatively intuitively, and in fact its recommended value range has been suggested. However, the density ratio is asymmetric for its center, and the possible error scale from its center, which should be close to the threshold, would depend on how the baseline policy is given. In order to maximize the values of regularization of policy, this paper proposes a new PPO derived using relative Pearson (RPE) divergence, therefore so-called PPO-RPE, to design the threshold adaptively. In PPO-RPE, the relative density ratio, which can be formed with symmetry, replaces the raw density ratio. Thanks to this symmetry, its error scale from center can easily be estimated, hence, the threshold can be adapted for the estimated error scale. From three simple benchmark simulations, the importance of algorithm-dependent threshold design is revealed. By simulating additional four locomotion tasks, it is verified that the proposed method statistically contributes to task accomplishment by appropriately restricting the policy updates.

Taisuke Kobayashi, Ryoma Watanuki

Extraction of low-dimensional latent space from high-dimensional observation data is essential to construct a real-time robot controller with a world model on the extracted latent space. However, there is no established method for tuning the dimension size of the latent space automatically, suffering from finding the necessary and sufficient dimension size, i.e. the minimal realization of the world model. In this study, we analyze and improve Tsallis-based variational autoencoder (q-VAE), and reveal that, under an appropriate configuration, it always facilitates making the latent space sparse. Even if the dimension size of the pre-specified latent space is redundant compared to the minimal realization, this sparsification collapses unnecessary dimensions, allowing for easy removal of them. We experimentally verified the benefits of the sparsification by the proposed method that it can easily find the necessary and sufficient six dimensions for a reaching task with a mobile manipulator that requires a six-dimensional state space. Moreover, by planning with such a minimal-realization world model learned in the extracted dimensions, the proposed method was able to exert a more optimal action sequence in real-time, reducing the reaching accomplishment time by around 20 %. The attached video is uploaded on youtube: https://youtu.be/-QjITrnxaRs

Taisuke Kobayashi

Robot control using reinforcement learning has become popular, but its learning process generally terminates halfway through an episode for safety and time-saving reasons. This study addresses the problem of the most popular exception handling that temporal-difference (TD) learning performs at such termination. That is, by forcibly assuming zero value after termination, unintentionally implicit underestimation or overestimation occurs, depending on the reward design in the normal states. When the episode is terminated due to task failure, the failure may be highly valued with the unintentional overestimation, and the wrong policy may be acquired. Although this problem can be avoided by paying attention to the reward design, it is essential in practical use of TD learning to review the exception handling at termination. This paper therefore proposes a method to intentionally underestimate the value after termination to avoid learning failures due to the unintentional overestimation. In addition, the degree of underestimation is adjusted according to the degree of stationarity at termination, thereby preventing excessive exploration due to the intentional underestimation. Simulations and real robot experiments showed that the proposed method can stably obtain the optimal policies for various tasks and reward designs. https://youtu.be/AxXr8uFOe7M

Taisuke Kobayashi

Soft actor-critic (SAC) in reinforcement learning is expected to be one of the next-generation robot control schemes. Its ability to maximize policy entropy would make a robotic controller robust to noise and perturbation, which is useful for real-world robot applications. However, the priority of maximizing the policy entropy is automatically tuned in the current implementation, the rule of which can be interpreted as one for equality constraint, binding the policy entropy into its specified lower bound. The current SAC is therefore no longer maximize the policy entropy, contrary to our expectation. To resolve this issue in SAC, this paper improves its implementation with a learnable state-dependent slack variable for appropriately handling the inequality constraint to maximize the policy entropy by reformulating it as the corresponding equality constraint. The introduced slack variable is optimized by a switching-type loss function that takes into account the dual objectives of satisfying the equality constraint and checking the lower bound. In Mujoco and Pybullet simulators, the modified SAC statistically achieved the higher robustness for adversarial attacks than before while regularizing the norm of action. A real-robot variable impedance task was demonstrated for showing the applicability of the modified SAC to real-world robot control. In particular, the modified SAC maintained adaptive behaviors for physical human-robot interaction, which had no experience at all during training. https://youtu.be/EH3xVtlVaJw

Taisuke Kobayashi

This paper introduces a novel method of adding intrinsic bonuses to task-oriented reward function in order to efficiently facilitate reinforcement learning search. While various bonuses have been designed to date, they are analogous to the depth-first and breadth-first search algorithms in graph theory. This paper, therefore, first designs two bonuses for each of them. Then, a heuristic gain scheduling is applied to the designed bonuses, inspired by the iterative deepening search, which is known to inherit the advantages of the two search algorithms. The proposed method is expected to allow agent to efficiently reach the best solution in deeper states by gradually exploring unknown states. In three locomotion tasks with dense rewards and three simple tasks with sparse rewards, it is shown that the two types of bonuses contribute to the performance improvement of the different tasks complementarily. In addition, by combining them with the proposed gain scheduling, all tasks can be accomplished with high performance.

Wendyam Eric Lionel Ilboudo, Taisuke Kobayashi, Kenji Sugimoto

Machine learning algorithms aim to find patterns from observations, which may include some noise, especially in robotics domain. To perform well even with such noise, we expect them to be able to detect outliers and discard them when needed. We therefore propose a new stochastic gradient optimization method, whose robustness is directly built in the algorithm, using the robust student-t distribution as its core idea. Adam, the popular optimization method, is modified with our method and the resultant optimizer, so-called TAdam, is shown to effectively outperform Adam in terms of robustness against noise on diverse task, ranging from regression and classification to reinforcement learning problems. The implementation of our algorithm can be found at https://github.com/Mahoumaru/TAdam.git

Taisuke Kobayashi, Takumi Aotani

This paper proposes a new design method for a stochastic control policy using a normalizing flow (NF). In reinforcement learning (RL), the policy is usually modeled as a distribution model with trainable parameters. When this parameterization has less expressiveness, it would fail to acquiring the optimal policy. A mixture model has capability of a universal approximation, but it with too much redundancy increases the computational cost, which can become a bottleneck when considering the use of real-time robot control. As another approach, NF, which is with additional parameters for invertible transformation from a simple stochastic model as a base, is expected to exert high expressiveness and lower computational cost. However, NF cannot compute its mean analytically due to complexity of the invertible transformation, and it lacks reliability because it retains stochastic behaviors after deployment for robot controller. This paper therefore designs a restricted NF (RNF) that achieves an analytic mean by appropriately restricting the invertible transformation. In addition, the expressiveness impaired by this restriction is regained using bimodal student-t distribution as its base, so-called Bit-RNF. In RL benchmarks, Bit-RNF policy outperformed the previous models. Finally, a real robot experiment demonstrated the applicability of Bit-RNF policy to real world. The attached video is uploaded on youtube: https://youtu.be/R_GJVZDW9bk

Taisuke Kobayashi, Shingo Murata

This paper proposes a novel stable learning theory for recurrent neural networks (RNNs), so-called variational adaptive noise and dropout (VAND). As stabilizing factors for RNNs, noise and dropout on the internal state of RNNs have been separately confirmed in previous studies. We reinterpret the optimization problem of RNNs as variational inference, showing that noise and dropout can be derived simultaneously by transforming the explicit regularization term arising in the optimization problem into implicit regularization. Their scale and ratio can also be adjusted appropriately to optimize the main objective of RNNs, respectively. In an imitation learning scenario with a mobile manipulator, only VAND is able to imitate sequential and periodic behaviors as instructed. https://youtu.be/UOho3Xr6A2w

Taisuke Kobayashi

Continual learning is the one of the most essential abilities for autonomous agents, which can incrementally learn daily-life skills. For this ultimate goal, a simple but powerful method, dark experience replay (DER), has been proposed recently. DER mitigates catastrophic forgetting, in which the skills acquired in the past are unintentionally forgotten, by stochastically storing the streaming data in a reservoir sampling (RS) buffer and by relearning them or retaining the past outputs for them. However, since DER considers multiple objectives, it will not function properly without appropriate weighting of them. In addition, the ability to retain past outputs inhibits learning if the past outputs are incorrect due to distribution shift or other effects. This is due to a tradeoff between memory consolidation and plasticity. The tradeoff is hidden even in the RS buffer, which gradually stops storing new data for new skills in it as data is continuously passed to it. To alleviate the tradeoff and achieve better balance, this paper proposes improvement strategies to each of DER and RS. Specifically, DER is improved with automatic adaptation of weights, block of replaying erroneous data, and correction of past outputs. RS is also improved with generalization of acceptance probability, stratification of plural buffers, and intentional omission of unnecessary data. These improvements are verified through multiple benchmarks including regression, classification, and reinforcement learning problems. As a result, the proposed methods achieve steady improvements in learning performance by balancing the memory consolidation and plasticity.

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.

Taisuke Kobayashi

Demand for deep reinforcement learning (DRL) is gradually increased to enable robots to perform complex tasks, while DRL is known to be unstable. As a technique to stabilize its learning, a target network that slowly and asymptotically matches a main network is widely employed to generate stable pseudo-supervised signals. Recently, T-soft update has been proposed as a noise-robust update rule for the target network and has contributed to improving the DRL performance. However, the noise robustness of T-soft update is specified by a hyperparameter, which should be tuned for each task, and is deteriorated by a simplified implementation. This study develops adaptive T-soft (AT-soft) update by utilizing the update rule in AdaTerm, which has been developed recently. In addition, the concern that the target network does not asymptotically match the main network is mitigated by a new consolidation for bringing the main network back to the target network. This so-called consolidated AT-soft (CAT-soft) update is verified through numerical simulations.

Taisuke Kobayashi

This paper proposes a new regularization technique for reinforcement learning (RL) towards making policy and value functions smooth and stable. RL is known for the instability of the learning process and the sensitivity of the acquired policy to noise. Several methods have been proposed to resolve these problems, and in summary, the smoothness of policy and value functions learned mainly in RL contributes to these problems. However, if these functions are extremely smooth, their expressiveness would be lost, resulting in not obtaining the global optimal solution. This paper therefore considers RL under local Lipschitz continuity constraint, so-called L2C2. By designing the spatio-temporal locally compact space for L2C2 from the state transition at each time step, the moderate smoothness can be achieved without loss of expressiveness. Numerical noisy simulations verified that the proposed L2C2 outperforms the task performance while smoothing out the robot action generated from the learned policy.

Taisuke Kobayashi, Takahito Enomoto

Autonomous driving has made great progress and been introduced in practical use step by step. On the other hand, the concept of personal mobility is also getting popular, and its autonomous driving specialized for individual drivers is expected for a new step. However, it is difficult to collect a large driving dataset, which is basically required for the learning of autonomous driving, from the individual driver of the personal mobility. In addition, when the driver is not familiar with the operation of the personal mobility, the dataset will contain non-optimal data. This study therefore focuses on an autonomous driving method for the personal mobility with such a small and noisy, so-called personal, dataset. Specifically, we introduce a new loss function based on Tsallis statistics that weights gradients depending on the original loss function and allows us to exclude noisy data in the optimization phase. In addition, we improve the visualization technique to verify whether the driver and the controller have the same region of interest. From the experimental results, we found that the conventional autonomous driving failed to drive properly due to the wrong operations in the personal dataset, and the region of interest was different from that of the driver. In contrast, the proposed method learned robustly against the errors and successfully drove automatically while paying attention to the similar region to the driver. Attached video is also uploaded on youtube: https://youtu.be/KEq8-bOxYQA

Wendyam Eric Lionel Ilboudo, Taisuke Kobayashi, Kenji Sugimoto

Behavioral cloning (BC) bears a high potential for safe and direct transfer of human skills to robots. However, demonstrations performed by human operators often contain noise or imperfect behaviors that can affect the efficiency of the imitator if left unchecked. In order to allow the imitators to effectively learn from imperfect demonstrations, we propose to employ the robust t-momentum optimization algorithm. This algorithm builds on the Student's t-distribution in order to deal with heavy-tailed data and reduce the effect of outlying observations. We extend the t-momentum algorithm to allow for an adaptive and automatic robustness and show empirically how the algorithm can be used to produce robust BC imitators against datasets with unknown heaviness. Indeed, the imitators trained with the t-momentum-based Adam optimizers displayed robustness to imperfect demonstrations on two different manipulation tasks with different robots and revealed the capability to take advantage of the additional data while reducing the adverse effect of non-optimal behaviors.

Taisuke Kobayashi, Shingo Murata, Tetsunari Inamura

This paper presents a new data-driven framework for analyzing periodic physical human-robot interaction (pHRI) in latent state space. To elaborate human understanding and/or robot control during pHRI, the model representing pHRI is critical. Recent developments of deep learning technologies would enable us to learn such a model from a dataset collected from the actual pHRI. Our framework is developed based on variational recurrent neural network (VRNN), which can inherently handle time-series data like one pHRI generates. This paper modifies VRNN in order to include the latent dynamics from robot to human explicitly. In addition, to analyze periodic motions like walking, we integrate a new recurrent network based on reservoir computing (RC), which has random and fixed connections between numerous neurons, with VRNN. By augmenting RC into complex domain, periodic behavior can be represented as the phase rotation in complex domain without decaying the amplitude. For verification of the proposed framework, a rope-rotation/swinging experiment was analyzed. The proposed framework, trained on the dataset collected from the experiment, achieved the latent state space where the differences in periodic motions can be distinguished. Such a well-distinguished space yielded the best prediction accuracy of the human observations and the robot actions. The attached video can be seen in youtube: https://youtu.be/umn0MVcIpsY

Taisuke Kobayashi

This paper proposes a new reinforcement learning with hyperbolic discounting. Combining a new temporal difference error with the hyperbolic discounting in recursive manner and reward-punishment framework, a new scheme to learn the optimal policy is derived. In simulations, it is found that the proposal outperforms the standard reinforcement learning, although the performance depends on the design of reward and punishment. In addition, the averages of discount factors w.r.t. reward and punishment are different from each other, like a sign effect in animal behaviors.

Taisuke Kobayashi

This paper addresses a new interpretation of the traditional optimization method in reinforcement learning (RL) as optimization problems using reverse Kullback-Leibler (KL) divergence, and derives a new optimization method using forward KL divergence, instead of reverse KL divergence in the optimization problems. Although RL originally aims to maximize return indirectly through optimization of policy, the recent work by Levine has proposed a different derivation process with explicit consideration of optimality as stochastic variable. This paper follows this concept and formulates the traditional learning laws for both value function and policy as the optimization problems with reverse KL divergence including optimality. Focusing on the asymmetry of KL divergence, the new optimization problems with forward KL divergence are derived. Remarkably, such new optimization problems can be regarded as optimistic RL. That optimism is intuitively specified by a hyperparameter converted from an uncertainty parameter. In addition, it can be enhanced when it is integrated with prioritized experience replay and eligibility traces, both of which accelerate learning. The effects of this expected optimism was investigated through learning tendencies on numerical simulations using Pybullet. As a result, moderate optimism accelerated learning and yielded higher rewards. In a realistic robotic simulation, the proposed method with the moderate optimism outperformed one of the state-of-the-art RL method.

Taisuke Kobayashi, Kenta Yoshizawa

Model-free or learning-based control, in particular, reinforcement learning (RL), is expected to be applied for complex robotic tasks. Traditional RL requires a policy to be optimized is state-dependent, that means, the policy is a kind of feedback (FB) controllers. Due to the necessity of correct state observation in such a FB controller, it is sensitive to sensing failures. To alleviate this drawback of the FB controllers, feedback error learning integrates one of them with a feedforward (FF) controller. RL can be improved by dealing with the FB/FF policies, but to the best of our knowledge, a methodology for learning them in a unified manner has not been developed. In this paper, we propose a new optimization problem for optimizing both the FB/FF policies simultaneously. Inspired by control as inference, the optimization problem considers minimization/maximization of divergences between trajectory, predicted by the composed policy and a stochastic dynamics model, and optimal/non-optimal trajectories. By approximating the stochastic dynamics model using variational method, we naturally derive a regularization between the FB/FF policies. In numerical simulations and a robot experiment, we verified that the proposed method can stably optimize the composed policy even with the different learning law from the traditional RL. In addition, we demonstrated that the FF policy is robust to the sensing failures and can hold the optimal motion. Attached video is also uploaded on youtube: https://youtu.be/zLL4uXIRmrE

Taisuke Kobayashi, Takanori Jin

To control humanoid robots, the reference pose of end effector(s) is planned in task space, then mapped into the reference joints by IK. By viewing that problem as approximate quadratic programming (QP), recent QP solvers can be applied to solve it precisely, but iterative numerical IK solvers based on Jacobian are still in high demand due to their low computational cost. However, the conventional Jacobian-based IK usually clamps the obtained joints during iteration according to the constraints in practice, causing numerical instability due to non-smoothed objective function. To alleviate the clamping problem, this study explicitly considers the joint constraints, especially the box constraints in this paper, inside the new IK solver. Specifically, instead of clamping, a mirror descent (MD) method with box-constrained real joint space and no-constrained mirror space is integrated with the Jacobian-based IK, so-called MD-IK. In addition, to escape local optima nearly on the boundaries of constraints, a heuristic technique, called $ε$-clamping, is implemented as margin in software level. Finally, to increase convergence speed, the acceleration method for MD is integrated assuming continuity of solutions at each time. As a result, the accelerated MD-IK achieved more stable and enough fast tracking performance compared to the conventional IK solvers. The low computational cost of the proposed method mitigated the time delay until the solution is obtained in real-time humanoid gait control, achieving a more stable gait.

Taisuke Kobayashi

The recent remarkable progress of deep reinforcement learning (DRL) stands on regularization of policy for stable and efficient learning. A popular method, named proximal policy optimization (PPO), has been introduced for this purpose. PPO clips density ratio of the latest and baseline policies with a threshold, while its minimization target is unclear. As another problem of PPO, the symmetric threshold is given numerically while the density ratio itself is in asymmetric domain, thereby causing unbalanced regularization of the policy. This paper therefore proposes a new variant of PPO by considering a regularization problem of relative Pearson (RPE) divergence, so-called PPO-RPE. This regularization yields the clear minimization target, which constrains the latest policy to the baseline one. Through its analysis, the intuitive threshold-based design consistent with the asymmetry of the threshold and the domain of density ratio can be derived. Through four benchmark tasks, PPO-RPE performed as well as or better than the conventional methods in terms of the task performance by the learned policy.

Taisuke Kobayashi, Wendyam Eric Lionel Ilboudo

This paper proposes a new robust update rule of target network for deep reinforcement learning (DRL), to replace the conventional update rule, given as an exponential moving average. The target network is for smoothly generating the reference signals for a main network in DRL, thereby reducing learning variance. The problem with its conventional update rule is the fact that all the parameters are smoothly copied with the same speed from the main network, even when some of them are trying to update toward the wrong directions. This behavior increases the risk of generating the wrong reference signals. Although slowing down the overall update speed is a naive way to mitigate wrong updates, it would decrease learning speed. To robustly update the parameters while keeping learning speed, a t-soft update method, which is inspired by student-t distribution, is derived with reference to the analogy between the exponential moving average and the normal distribution. Through the analysis of the derived t-soft update, we show that it takes over the properties of the student-t distribution. Specifically, with a heavy-tailed property of the student-t distribution, the t-soft update automatically excludes extreme updates that differ from past experiences. In addition, when the updates are similar to the past experiences, it can mitigate the learning delay by increasing the amount of updates. In PyBullet robotics simulations for DRL, an online actor-critic algorithm with the t-soft update outperformed the conventional methods in terms of the obtained return and/or its variance. From the training process by the t-soft update, we found that the t-soft update is globally consistent with the standard soft update, and the update rates are locally adjusted for acceleration or suppression.

Taisuke Kobayashi

Deep reinforcement learning (DRL) is one promising approach to teaching robots to perform complex tasks. Because methods that directly reuse the stored experience data cannot follow the change of the environment in robotic problems with a time-varying environment, online DRL is required. The eligibility traces method is well known as an online learning technique for improving sample efficiency in traditional reinforcement learning with linear regressors rather than DRL. The dependency between parameters of deep neural networks would destroy the eligibility traces, which is why they are not integrated with DRL. Although replacing the gradient with the most influential one rather than accumulating the gradients as the eligibility traces can alleviate this problem, the replacing operation reduces the number of reuses of previous experiences. To address these issues, this study proposes a new eligibility traces method that can be used even in DRL while maintaining high sample efficiency. When the accumulated gradients differ from those computed using the latest parameters, the proposed method takes into account the divergence between the past and latest parameters to adaptively decay the eligibility traces. Bregman divergences between outputs computed by the past and latest parameters are exploited due to the infeasible computational cost of the divergence between the past and latest parameters. In addition, a generalized method with multiple time-scale traces is designed for the first time. This design allows for the replacement of the most influential adaptively accumulated (decayed) eligibility traces.