Surgical robot task automation has recently attracted great attention due to its potential to benefit both surgeons and patients. Reinforcement learning (RL) based approaches have demonstrated promising ability to provide solutions to automated surgical manipulations on various tasks. To address the exploration challenge, expert demonstrations can be utilized to enhance the learning efficiency via imitation learning (IL) approaches. However, the successes of such methods normally rely on both states and action labels. Unfortunately action labels can be hard to capture or their manual annotation is prohibitively expensive owing to the requirement for expert knowledge. It therefore remains an appealing and open problem to leverage expert demonstrations composed of pure states in RL. In this work, we present an actor-critic RL framework, termed AC-SSIL, to overcome this challenge of learning with state-only demonstrations collected by following an unknown expert policy. It adopts a self-supervised IL method, dubbed SSIL, to effectively incorporate demonstrated states into RL paradigms by retrieving from demonstrates the nearest neighbours of the query state and utilizing the bootstrapping of actor networks. We showcase through experiments on an open-source surgical simulation platform that our method delivers remarkable improvements over the RL baseline and exhibits comparable performance against action based IL methods, which implies the efficacy and potential of our method for expert demonstration-guided learning scenarios.

Specifying tasks for robotic systems traditionally requires coding expertise, deep domain knowledge, and significant time investment. While learning from demonstration offers a promising alternative, existing methods often struggle with tasks of longer horizons. To address this limitation, we introduce a computationally efficient approach for learning probabilistic deterministic finite automata (PDFA) that capture task structures and expert preferences directly from demonstrations. Our approach infers sub-goals and their temporal dependencies, producing an interpretable task specification that domain experts can easily understand and adjust. We validate our method through experiments involving object manipulation tasks, showcasing how our method enables a robot arm to effectively replicate diverse expert strategies while adapting to changing conditions.

The emergence of mathematical concepts, such as number systems, is an understudied area in AI for mathematics and reasoning. It has previously been shown Carlsson et al. (2021) that by using reinforcement learning (RL), agents can derive simple approximate and exact-restricted numeral systems. However, it is a major challenge to show how more complex recursive numeral systems, similar to the one utilised in English, could arise via a simple learning mechanism such as RL. Here, we introduce an approach towards deriving a mechanistic explanation of the emergence of recursive number systems where we consider an RL agent which directly optimizes a lexicon under a given meta-grammar. Utilising a slightly modified version of the seminal meta-grammar of Hurford (1975), we demonstrate that our RL agent can effectively modify the lexicon towards Pareto-optimal configurations which are comparable to those observed within human numeral systems.

Modern listwise recommendation systems need to consider both long-term user perceptions and short-term interest shifts. Reinforcement learning can be applied on recommendation to study such a problem but is also subject to large search space, sparse user feedback and long interactive latency. Motivated by recent progress in hierarchical reinforcement learning, we propose a novel framework called mccHRL to provide different levels of temporal abstraction on listwise recommendation. Within the hierarchical framework, the high-level agent studies the evolution of user perception, while the low-level agent produces the item selection policy by modeling the process as a sequential decision-making problem. We argue that such framework has a well-defined decomposition of the outra-session context and the intra-session context, which are encoded by the high-level and low-level agents, respectively. To verify this argument, we implement both a simulator-based environment and an industrial dataset-based experiment. Results observe significant performance improvement by our method, compared with several well-known baselines. Data and codes have been made public.

Pedipulation leverages the feet of legged robots for mobile manipulation, eliminating the need for dedicated robotic arms. While previous works have showcased blind and task-specific pedipulation skills, they fail to account for static and dynamic obstacles in the environment. To address this limitation, we introduce a reinforcement learning-based approach to train a whole-body obstacle-aware policy that tracks foot position commands while simultaneously avoiding obstacles. Despite training the policy in only five different static scenarios in simulation, we show that it generalizes to unknown environments with different numbers and types of obstacles. We analyze the performance of our method through a set of simulation experiments and successfully deploy the learned policy on the ANYmal quadruped, demonstrating its capability to follow foot commands while navigating around static and dynamic obstacles.

Wirelessly streaming high quality 360 degree videos is still a challenging problem. When there are many users watching different 360 degree videos and competing for the computing and communication resources, the streaming algorithm at hand should maximize the average quality of experience (QoE) while guaranteeing a minimum rate for each user. In this paper, we propose a cross layer optimization approach that maximizes the available rate to each user and efficiently uses it to maximize users' QoE. Particularly, we consider a tile based 360 degree video streaming, and we optimize a QoE metric that balances the tradeoff between maximizing each user's QoE and ensuring fairness among users. We show that the problem can be decoupled into two interrelated subproblems: (i) a physical layer subproblem whose objective is to find the download rate for each user, and (ii) an application layer subproblem whose objective is to use that rate to find a quality decision per tile such that the user's QoE is maximized. We prove that the physical layer subproblem can be solved optimally with low complexity and an actor-critic deep reinforcement learning (DRL) is proposed to leverage the parallel training of multiple independent agents and solve the application layer subproblem. Extensive experiments reveal the robustness of our scheme and demonstrate its significant performance improvement compared to several baseline algorithms.

Recently, the pre-training of decision transformers (DT) using a different domain, such as natural language text, has generated significant attention in offline reinforcement learning (Offline RL). Although this cross-domain pre-training approach achieves superior performance compared to training from scratch in environments required short-term planning ability, the mechanisms by which pre-training benefits the fine-tuning phase remain unclear. Furthermore, we point out that the cross-domain pre-training approach hinders the extraction of distant information in environments like PointMaze that require long-term planning ability, leading to performance that is much worse than training DT from scratch. This work first analyzes these issues and found that Markov Matrix, a component that exists in pre-trained attention heads, is the key to explain the significant performance disparity of pre-trained models in different planning abilities. Inspired by our analysis, we propose a general method GPT-DTMA, which equips a pre-trained DT with Mixture of Attention (MoA), to enable adaptive learning and accommodating diverse attention requirements during fine-tuning. Extensive experiments demonstrate that the effectiveness of GPT-DTMA: it achieves superior performance in short-term environments compared to baselines, and in long-term environments, it mitigates the negative impact caused by Markov Matrix, achieving results comparable to those of DT trained from scratch.

World models are increasingly pivotal in interpreting and simulating the rules and actions of complex environments. Genie, a recent model, excels at learning from visually diverse environments but relies on costly human-collected data. We observe that their alternative method of using random agents is too limited to explore the environment. We propose to improve the model by employing reinforcement learning based agents for data generation. This approach produces diverse datasets that enhance the model's ability to adapt and perform well across various scenarios and realistic actions within the environment. In this paper, we first release the model GenieRedux - an implementation based on Genie. Additionally, we introduce GenieRedux-G, a variant that uses the agent's readily available actions to factor out action prediction uncertainty during validation. Our evaluation, including a replication of the Coinrun case study, shows that GenieRedux-G achieves superior visual fidelity and controllability using the trained agent exploration. The proposed approach is reproducable, scalable and adaptable to new types of environments. Our codebase is available at https://github.com/insait-institute/GenieRedux .

Q-learning is a widely used algorithm in reinforcement learning (RL), but its convergence can be slow, especially when the discount factor is close to one. Successive Over-Relaxation (SOR) Q-learning, which introduces a relaxation factor to speed up convergence, addresses this issue but has two major limitations: In the tabular setting, the relaxation parameter depends on transition probability, making it not entirely model-free, and it suffers from overestimation bias. To overcome these limitations, we propose a sample-based, model-free double SOR Q-learning algorithm. Theoretically and empirically, this algorithm is shown to be less biased than SOR Q-learning. Further, in the tabular setting, the convergence analysis under boundedness assumptions on iterates is discussed. The proposed algorithm is extended to large-scale problems using deep RL. Finally, the tabular version of the proposed algorithm is compared using roulette and grid world environments, while the deep RL version is tested on a maximization bias example and OpenAI Gym environments.

Deep Reinforcement Learning techniques are achieving state-of-the-art results in robust legged locomotion. While there exists a wide variety of legged platforms such as quadruped, humanoids, and hexapods, the field is still missing a single learning framework that can control all these different embodiments easily and effectively and possibly transfer, zero or few-shot, to unseen robot embodiments. We introduce URMA, the Unified Robot Morphology Architecture, to close this gap. Our framework brings the end-to-end Multi-Task Reinforcement Learning approach to the realm of legged robots, enabling the learned policy to control any type of robot morphology. The key idea of our method is to allow the network to learn an abstract locomotion controller that can be seamlessly shared between embodiments thanks to our morphology-agnostic encoders and decoders. This flexible architecture can be seen as a potential first step in building a foundation model for legged robot locomotion. Our experiments show that URMA can learn a locomotion policy on multiple embodiments that can be easily transferred to unseen robot platforms in simulation and the real world.