Fabian Otto

Fabian Otto, Onur Celik, Hongyi Zhou et al.

\Episode-based reinforcement learning (ERL) algorithms treat reinforcement learning (RL) as a black-box optimization problem where we learn to select a parameter vector of a controller, often represented as a movement primitive, for a given task descriptor called a context. ERL offers several distinct benefits in comparison to step-based RL. It generates smooth control trajectories, can handle non-Markovian reward definitions, and the resulting exploration in parameter space is well suited for solving sparse reward settings. Yet, the high dimensionality of the movement primitive parameters has so far hampered the effective use of deep RL methods. In this paper, we present a new algorithm for deep ERL. It is based on differentiable trust region layers, a successful on-policy deep RL algorithm. These layers allow us to specify trust regions for the policy update that are solved exactly for each state using convex optimization, which enables policies learning with the high precision required for the ERL. We compare our ERL algorithm to state-of-the-art step-based algorithms in many complex simulated robotic control tasks. In doing so, we investigate different reward formulations - dense, sparse, and non-Markovian. While step-based algorithms perform well only on dense rewards, ERL performs favorably on sparse and non-Markovian rewards. Moreover, our results show that the sparse and the non-Markovian rewards are also often better suited to define the desired behavior, allowing us to obtain considerably higher quality policies compared to step-based RL.

Fabian Otto, Hongyi Zhou, Onur Celik et al.

We introduce a novel deep reinforcement learning (RL) approach called Movement Primitive-based Planning Policy (MP3). By integrating movement primitives (MPs) into the deep RL framework, MP3 enables the generation of smooth trajectories throughout the whole learning process while effectively learning from sparse and non-Markovian rewards. Additionally, MP3 maintains the capability to adapt to changes in the environment during execution. Although many early successes in robot RL have been achieved by combining RL with MPs, these approaches are often limited to learning single stroke-based motions, lacking the ability to adapt to task variations or adjust motions during execution. Building upon our previous work, which introduced an episode-based RL method for the non-linear adaptation of MP parameters to different task variations, this paper extends the approach to incorporating replanning strategies. This allows adaptation of the MP parameters throughout motion execution, addressing the lack of online motion adaptation in stochastic domains requiring feedback. We compared our approach against state-of-the-art deep RL and RL with MPs methods. The results demonstrated improved performance in sophisticated, sparse reward settings and in domains requiring replanning.

Ge Li, Zeqi Jin, Michael Volpp et al.

Movement Primitives (MPs) are a well-known concept to represent and generate modular trajectories. MPs can be broadly categorized into two types: (a) dynamics-based approaches that generate smooth trajectories from any initial state, e. g., Dynamic Movement Primitives (DMPs), and (b) probabilistic approaches that capture higher-order statistics of the motion, e. g., Probabilistic Movement Primitives (ProMPs). To date, however, there is no method that unifies both, i. e. that can generate smooth trajectories from an arbitrary initial state while capturing higher-order statistics. In this paper, we introduce a unified perspective of both approaches by solving the ODE underlying the DMPs. We convert expensive online numerical integration of DMPs into basis functions that can be computed offline. These basis functions can be used to represent trajectories or trajectory distributions similar to ProMPs while maintaining all the properties of dynamical systems. Since we inherit the properties of both methodologies, we call our proposed model Probabilistic Dynamic Movement Primitives (ProDMPs). Additionally, we embed ProDMPs in deep neural network architecture and propose a new cost function for efficient end-to-end learning of higher-order trajectory statistics. To this end, we leverage Bayesian Aggregation for non-linear iterative conditioning on sensory inputs. Our proposed model achieves smooth trajectory generation, goal-attractor convergence, correlation analysis, non-linear conditioning, and online re-planing in one framework.

Philipp Becker, Sebastian Mossburger, Fabian Otto et al.

Learning self-supervised representations using reconstruction or contrastive losses improves performance and sample complexity of image-based and multimodal reinforcement learning (RL). Here, different self-supervised loss functions have distinct advantages and limitations depending on the information density of the underlying sensor modality. Reconstruction provides strong learning signals but is susceptible to distractions and spurious information. While contrastive approaches can ignore those, they may fail to capture all relevant details and can lead to representation collapse. For multimodal RL, this suggests that different modalities should be treated differently based on the amount of distractions in the signal. We propose Contrastive Reconstructive Aggregated representation Learning (CoRAL), a unified framework enabling us to choose the most appropriate self-supervised loss for each sensor modality and allowing the representation to better focus on relevant aspects. We evaluate CoRAL's benefits on a wide range of tasks with images containing distractions or occlusions, a new locomotion suite, and a challenging manipulation suite with visually realistic distractions. Our results show that learning a multimodal representation by combining contrastive and reconstruction-based losses can significantly improve performance and solve tasks that are out of reach for more naive representation learning approaches and other recent baselines.

Fabian Otto, Philipp Becker, Ngo Anh Vien et al.

Trust region methods are a popular tool in reinforcement learning as they yield robust policy updates in continuous and discrete action spaces. However, enforcing such trust regions in deep reinforcement learning is difficult. Hence, many approaches, such as Trust Region Policy Optimization (TRPO) and Proximal Policy Optimization (PPO), are based on approximations. Due to those approximations, they violate the constraints or fail to find the optimal solution within the trust region. Moreover, they are difficult to implement, often lack sufficient exploration, and have been shown to depend on seemingly unrelated implementation choices. In this work, we propose differentiable neural network layers to enforce trust regions for deep Gaussian policies via closed-form projections. Unlike existing methods, those layers formalize trust regions for each state individually and can complement existing reinforcement learning algorithms. We derive trust region projections based on the Kullback-Leibler divergence, the Wasserstein L2 distance, and the Frobenius norm for Gaussian distributions. We empirically demonstrate that those projection layers achieve similar or better results than existing methods while being almost agnostic to specific implementation choices. The code is available at https://git.io/Jthb0.

Fabian Otto, Philipp Becker, Ngo Anh Vien et al.

Existing off-policy reinforcement learning algorithms often rely on an explicit state-action-value function representation, which can be problematic in high-dimensional action spaces due to the curse of dimensionality. This reliance results in data inefficiency as maintaining a state-action-value function in such spaces is challenging. We present an efficient approach that utilizes only a state-value function as the critic for off-policy deep reinforcement learning. This approach, which we refer to as Vlearn, effectively circumvents the limitations of existing methods by eliminating the necessity for an explicit state-action-value function. To this end, we leverage a weighted importance sampling loss for learning deep value functions from off-policy data. While this is common for linear methods, it has not been combined with deep value function networks. This transfer to deep methods is not straightforward and requires novel design choices such as robust policy updates, twin value function networks to avoid an optimization bias, and importance weight clipping. We also present a novel analysis of the variance of our estimate compared to commonly used importance sampling estimators such as V-trace. Our approach improves sample complexity as well as final performance and ensures consistent and robust performance across various benchmark tasks. Eliminating the state-action-value function in Vlearn facilitates a streamlined learning process, yielding high-return agents.

Philipp Becker, Niklas Freymuth, Serge Thilges et al.

On-policy Reinforcement Learning (RL) with PPO-like clip objectives has become the standard choice for reward-based fine-tuning of large language models (LLMs). Although recent work has explored improved estimators of advantages and normalization, the clipping mechanism itself has remained untouched. Originally introduced as a proxy for principled KL-based trust regions, clipping is a crude approximation that often causes unstable updates and suboptimal performance. We replace the clip objective with a novel discrete differentiable trust region projection, which provides principled token-level KL constraints. The projection operates on a sparse subset of the model's most important token logits to balance computational cost and projection effectiveness. Our approach, Trust Region Optimization for Large Language Models (TROLL), serves as a direct replacement for PPO-like clipping during training and does not alter the model's inference behavior. Across datasets, model families, and advantage-estimation methods, TROLL consistently outperforms PPO-like clipping in terms of training speed, stability, and final success rates.

Ge Li, Hongyi Zhou, Dominik Roth et al.

Current advancements in reinforcement learning (RL) have predominantly focused on learning step-based policies that generate actions for each perceived state. While these methods efficiently leverage step information from environmental interaction, they often ignore the temporal correlation between actions, resulting in inefficient exploration and unsmooth trajectories that are challenging to implement on real hardware. Episodic RL (ERL) seeks to overcome these challenges by exploring in parameters space that capture the correlation of actions. However, these approaches typically compromise data efficiency, as they treat trajectories as opaque \emph{black boxes}. In this work, we introduce a novel ERL algorithm, Temporally-Correlated Episodic RL (TCE), which effectively utilizes step information in episodic policy updates, opening the 'black box' in existing ERL methods while retaining the smooth and consistent exploration in parameter space. TCE synergistically combines the advantages of step-based and episodic RL, achieving comparable performance to recent ERL methods while maintaining data efficiency akin to state-of-the-art (SoTA) step-based RL.