Scaling CrossQ with Weight Normalization

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics which are emphasized by higher UTDs, particularly Q-bias explosion and the growing magnitude of critic network weights. To address this, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, prevents potential loss of plasticity and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive or superior performance across a range of challenging tasks on the DeepMind control benchmark, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

Towards Safe Robot Foundation Models Using Inductive Biases

Safety is a critical requirement for the real-world deployment of robotic systems. Unfortunately, while current robot foundation models show promising generalization capabilities across a wide variety of tasks, they fail to address safety, an important aspect for ensuring long-term operation. Current robot foundation models assume that safe behavior should emerge by learning from a sufficiently large dataset of demonstrations. However, this approach has two clear major drawbacks. Firstly, there are no formal safety guarantees for a behavior cloning policy trained using supervised learning. Secondly, without explicit knowledge of any safety constraints, the policy may require an unreasonable number of additional demonstrations to even approximate the desired constrained behavior. To solve these key issues, we show how we can instead combine robot foundation models with geometric inductive biases using ATACOM, a safety layer placed after the foundation policy that ensures safe state transitions by enforcing action constraints. With this approach, we can ensure formal safety guarantees for generalist policies without providing extensive demonstrations of safe behavior, and without requiring any specific fine-tuning for safety. Our experiments show that our approach can be beneficial both for classical manipulation tasks, where we avoid unwanted collisions with irrelevant objects, and for dynamic tasks, such as the robot air hockey environment, where we can generate fast trajectories respecting complex tasks and joint space constraints.

Gait in Eight: Efficient On-Robot Learning for Omnidirectional Quadruped Locomotion

On-robot Reinforcement Learning is a promising approach to train embodiment-aware policies for legged robots. However, the computational constraints of real-time learning on robots pose a significant challenge. We present a framework for efficiently learning quadruped locomotion in just 8 minutes of raw real-time training utilizing the sample efficiency and minimal computational overhead of the new off-policy algorithm CrossQ. We investigate two control architectures: Predicting joint target positions for agile, high-speed locomotion and Central Pattern Generators for stable, natural gaits. While prior work focused on learning simple forward gaits, our framework extends on-robot learning to omnidirectional locomotion. We demonstrate the robustness of our approach in different indoor and outdoor environments.

Towards Safe Robot Foundation Models

Robot foundation models hold the potential for deployment across diverse environments, from industrial applications to household tasks. While current research focuses primarily on the policies' generalization capabilities across a variety of tasks, it fails to address safety, a critical requirement for deployment on real-world systems. In this paper, we introduce a safety layer designed to constrain the action space of any generalist policy appropriately. Our approach uses ATACOM, a safe reinforcement learning algorithm that creates a safe action space and, therefore, ensures safe state transitions. By extending ATACOM to generalist policies, our method facilitates their deployment in safety-critical scenarios without requiring any specific safety fine-tuning. We demonstrate the effectiveness of this safety layer in an air hockey environment, where it prevents a puck-hitting agent from colliding with its surroundings, a failure observed in generalist policies.

Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics, which are emphasized by higher UTD ratios. To address these, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, has been shown to prevent potential loss of plasticity and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive performance across 25 challenging continuous control tasks on the DeepMind Control Suite and Myosuite benchmarks, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a simple yet robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

Analysing the Interplay of Vision and Touch for Dexterous Insertion Tasks

Robotic insertion tasks remain challenging due to uncertainties in perception and the need for precise control, particularly in unstructured environments. While humans seamlessly combine vision and touch for such tasks, effectively integrating these modalities in robotic systems is still an open problem. Our work presents an extensive analysis of the interplay between visual and tactile feedback during dexterous insertion tasks, showing that tactile sensing can greatly enhance success rates on challenging insertions with tight tolerances and varied hole orientations that vision alone cannot solve. These findings provide valuable insights for designing more effective multi-modal robotic control systems and highlight the critical role of tactile feedback in contact-rich manipulation tasks.

Learning Tactile Insertion in the Real World

Humans have exceptional tactile sensing capabilities, which they can leverage to solve challenging, partially observable tasks that cannot be solved from visual observation alone. Research in tactile sensing attempts to unlock this new input modality for robots. Lately, these sensors have become cheaper and, thus, widely available. At the same time, the question of how to integrate them into control loops is still an active area of research, with central challenges being partial observability and the contact-rich nature of manipulation tasks. In this study, we propose to use Reinforcement Learning to learn an end-to-end policy, mapping directly from tactile sensor readings to actions. Specifically, we use Dreamer-v3 on a challenging, partially observable robotic insertion task with a Franka Research 3, both in simulation and on a real system. For the real setup, we built a robotic platform capable of resetting itself fully autonomously, allowing for extensive training runs without human supervision. Our preliminary results indicate that Dreamer is capable of utilizing tactile inputs to solve robotic manipulation tasks in simulation and reality. Furthermore, we find that providing the robot with tactile feedback generally improves task performance, though, in our setup, we do not yet include other sensing modalities. In the future, we plan to utilize our platform to evaluate a wide range of other Reinforcement Learning algorithms on tactile tasks.

Iterated $Q$-Network: Beyond the One-Step Bellman Operator

Value-based Reinforcement Learning (RL) methods rely on the application of the Bellman operator, which needs to be approximated from samples. Most approaches consist of an iterative scheme alternating the application of the Bellman operator and a subsequent projection step onto a considered function space. However, we observe that these algorithms can be improved by considering multiple iterations of the Bellman operator at once. Thus, we introduce iterated $Q$-Networks (iQN), a novel approach that learns a sequence of $Q$-function approximations where each $Q$-function serves as the target for the next one in a chain of consecutive Bellman iterations. We demonstrate that iQN is theoretically sound and show how it can be seamlessly used in value-based and actor-critic methods. We empirically demonstrate its advantages on Atari $2600$ games and in continuous-control MuJoCo environments.

Pseudo-Likelihood Inference

Simulation-Based Inference (SBI) is a common name for an emerging family of approaches that infer the model parameters when the likelihood is intractable. Existing SBI methods either approximate the likelihood, such as Approximate Bayesian Computation (ABC) or directly model the posterior, such as Sequential Neural Posterior Estimation (SNPE). While ABC is efficient on low-dimensional problems, on higher-dimensional tasks, it is generally outperformed by SNPE, which leverages function approximation. In this paper, we propose Pseudo-Likelihood Inference (PLI), a new method that brings neural approximation into ABC, making it competitive on challenging Bayesian system identification tasks. By utilizing integral probability metrics, we introduce a smooth likelihood kernel with an adaptive bandwidth that is updated based on information-theoretic trust regions. Thanks to this formulation, our method (i) allows for optimizing neural posteriors via gradient descent, (ii) does not rely on summary statistics, and (iii) enables multiple observations as input. In comparison to SNPE, it leads to improved performance when more data is available. The effectiveness of PLI is evaluated on four classical SBI benchmark tasks and on a highly dynamic physical system, showing particular advantages on stochastic simulations and multi-modal posterior landscapes.

Diminishing Return of Value Expansion Methods in Model-Based Reinforcement Learning

Model-based reinforcement learning is one approach to increase sample efficiency. However, the accuracy of the dynamics model and the resulting compounding error over modelled trajectories are commonly regarded as key limitations. A natural question to ask is: How much more sample efficiency can be gained by improving the learned dynamics models? Our paper empirically answers this question for the class of model-based value expansion methods in continuous control problems. Value expansion methods should benefit from increased model accuracy by enabling longer rollout horizons and better value function approximations. Our empirical study, which leverages oracle dynamics models to avoid compounding model errors, shows that (1) longer horizons increase sample efficiency, but the gain in improvement decreases with each additional expansion step, and (2) the increased model accuracy only marginally increases the sample efficiency compared to learned models with identical horizons. Therefore, longer horizons and increased model accuracy yield diminishing returns in terms of sample efficiency. These improvements in sample efficiency are particularly disappointing when compared to model-free value expansion methods. Even though they introduce no computational overhead, we find their performance to be on-par with model-based value expansion methods. Therefore, we conclude that the limitation of model-based value expansion methods is not the model accuracy of the learned models. While higher model accuracy is beneficial, our experiments show that even a perfect model will not provide an un-rivalled sample efficiency but that the bottleneck lies elsewhere.