Olivier Sigaud

Thomas Carta, Clément Romac, Thomas Wolf et al.

Recent works successfully leveraged Large Language Models' (LLM) abilities to capture abstract knowledge about world's physics to solve decision-making problems. Yet, the alignment between LLMs' knowledge and the environment can be wrong and limit functional competence due to lack of grounding. In this paper, we study an approach (named GLAM) to achieve this alignment through functional grounding: we consider an agent using an LLM as a policy that is progressively updated as the agent interacts with the environment, leveraging online Reinforcement Learning to improve its performance to solve goals. Using an interactive textual environment designed to study higher-level forms of functional grounding, and a set of spatial and navigation tasks, we study several scientific questions: 1) Can LLMs boost sample efficiency for online learning of various RL tasks? 2) How can it boost different forms of generalization? 3) What is the impact of online learning? We study these questions by functionally grounding several variants (size, architecture) of FLAN-T5.

Hugo Caselles-Dupré, Olivier Sigaud, Mohamed Chetouani

Learning from demonstration methods usually leverage close to optimal demonstrations to accelerate training. By contrast, when demonstrating a task, human teachers deviate from optimal demonstrations and pedagogically modify their behavior by giving demonstrations that best disambiguate the goal they want to demonstrate. Analogously, human learners excel at pragmatically inferring the intent of the teacher, facilitating communication between the two agents. These mechanisms are critical in the few demonstrations regime, where inferring the goal is more difficult. In this paper, we implement pedagogy and pragmatism mechanisms by leveraging a Bayesian model of Goal Inference from demonstrations (BGI). We highlight the benefits of this model in multi-goal teacher-learner setups with two artificial agents that learn with goal-conditioned Reinforcement Learning. We show that combining BGI-agents (a pedagogical teacher and a pragmatic learner) results in faster learning and reduced goal ambiguity over standard learning from demonstrations, especially in the few demonstrations regime. We provide the code for our experiments (https://github.com/Caselles/NeurIPS22-demonstrations-pedagogy-pragmatism), as well as an illustrative video explaining our approach (https://youtu.be/V4n16IjkNyw).

Thomas Carta, Pierre-Yves Oudeyer, Olivier Sigaud et al.

Reinforcement learning (RL) in long horizon and sparse reward tasks is notoriously difficult and requires a lot of training steps. A standard solution to speed up the process is to leverage additional reward signals, shaping it to better guide the learning process. In the context of language-conditioned RL, the abstraction and generalisation properties of the language input provide opportunities for more efficient ways of shaping the reward. In this paper, we leverage this idea and propose an automated reward shaping method where the agent extracts auxiliary objectives from the general language goal. These auxiliary objectives use a question generation (QG) and question answering (QA) system: they consist of questions leading the agent to try to reconstruct partial information about the global goal using its own trajectory. When it succeeds, it receives an intrinsic reward proportional to its confidence in its answer. This incentivizes the agent to generate trajectories which unambiguously explain various aspects of the general language goal. Our experimental study shows that this approach, which does not require engineer intervention to design the auxiliary objectives, improves sample efficiency by effectively directing exploration.

Olivier Sigaud

Deep neuroevolution and deep Reinforcement Learning have received a lot of attention in the last years. Some works have compared them, highlighting theirs pros and cons, but an emerging trend consists in combining them so as to benefit from the best of both worlds. In this paper, we provide a survey of this emerging trend by organizing the literature into related groups of works and casting all the existing combinations in each group into a generic framework. We systematically cover all easily available papers irrespective of their publication status, focusing on the combination mechanisms rather than on the experimental results. In total, we cover 45 algorithms more recent than 2017. We hope this effort will favor the growth of the domain by facilitating the understanding of the relationships between the methods, leading to deeper analyses, outlining missing useful comparisons and suggesting new combinations of mechanisms.

Maël Franceschetti, Coline Lacoux, Ryan Ohouens et al.

The SWIMMER environment is a standard benchmark in reinforcement learning (RL). In particular, it is often used in papers comparing or combining RL methods with direct policy search methods such as genetic algorithms or evolution strategies. A lot of these papers report poor performance on SWIMMER from RL methods and much better performance from direct policy search methods. In this technical report we show that the low performance of RL methods on SWIMMER simply comes from the inadequate tuning of an important hyper-parameter, the discount factor. Furthermore we show that, by setting this hyper-parameter to a correct value, the issue can be easily fixed. Finally, for a set of often used RL algorithms, we provide a set of successful hyper-parameters obtained with the Stable Baselines3 library and its RL Zoo.

Olivier Sigaud, Gianluca Baldassarre, Cedric Colas et al.

A lot of recent machine learning research papers have ``open-ended learning'' in their title. But very few of them attempt to define what they mean when using the term. Even worse, when looking more closely there seems to be no consensus on what distinguishes open-ended learning from related concepts such as continual learning, lifelong learning or autotelic learning. In this paper, we contribute to fixing this situation. After illustrating the genealogy of the concept and more recent perspectives about what it truly means, we outline that open-ended learning is generally conceived as a composite notion encompassing a set of diverse properties. In contrast with previous approaches, we propose to isolate a key elementary property of open-ended processes, which is to produce elements from time to time (e.g., observations, options, reward functions, and goals), over an infinite horizon, that are considered novel from an observer's perspective. From there, we build the notion of open-ended learning problems and focus in particular on the subset of open-ended goal-conditioned reinforcement learning problems in which agents can learn a growing repertoire of goal-driven skills. Finally, we highlight the work that remains to be performed to fill the gap between our elementary definition and the more involved notions of open-ended learning that developmental AI researchers may have in mind.

Hugo Caselles-Dupré, Olivier Sigaud, Mohamed Chetouani

Teaching an agent to perform new tasks using natural language can easily be hindered by ambiguities in interpretation. When a teacher provides an instruction to a learner about an object by referring to its features, the learner can misunderstand the teacher's intentions, for instance if the instruction ambiguously refer to features of the object, a phenomenon called referential ambiguity. We study how two concepts derived from cognitive sciences can help resolve those referential ambiguities: pedagogy (selecting the right instructions) and pragmatism (learning the preferences of the other agents using inductive reasoning). We apply those ideas to a teacher/learner setup with two artificial agents on a simulated robotic task (block-stacking). We show that these concepts improve sample efficiency for training the learner.

Clémence Grislain, Hugo Caselles-Dupré, Olivier Sigaud et al.

Good teachers always tailor their explanations to the learners. Cognitive scientists model this process under the rationality principle: teachers try to maximise the learner's utility while minimising teaching costs. To this end, human teachers seem to build mental models of the learner's internal state, a capacity known as Theory of Mind (ToM). Inspired by cognitive science, we build on Bayesian ToM mechanisms to design teacher agents that, like humans, tailor their teaching strategies to the learners. Our ToM-equipped teachers construct models of learners' internal states from observations and leverage them to select demonstrations that maximise the learners' rewards while minimising teaching costs. Our experiments in simulated environments demonstrate that learners taught this way are more efficient than those taught in a learner-agnostic way. This effect gets stronger when the teacher's model of the learner better aligns with the actual learner's state, either using a more accurate prior or after accumulating observations of the learner's behaviour. This work is a first step towards social machines that teach us and each other, see https://teacher-with-tom.github.io.

Nicolas Castanet, Sylvain Lamprier, Olivier Sigaud

In multi-goal Reinforcement Learning, an agent can share experience between related training tasks, resulting in better generalization for new tasks at test time. However, when the goal space has discontinuities and the reward is sparse, a majority of goals are difficult to reach. In this context, a curriculum over goals helps agents learn by adapting training tasks to their current capabilities. In this work we propose Stein Variational Goal Generation (SVGG), which samples goals of intermediate difficulty for the agent, by leveraging a learned predictive model of its goal reaching capabilities. The distribution of goals is modeled with particles that are attracted in areas of appropriate difficulty using Stein Variational Gradient Descent. We show that SVGG outperforms state-of-the-art multi-goal Reinforcement Learning methods in terms of success coverage in hard exploration problems, and demonstrate that it is endowed with a useful recovery property when the environment changes.

Alexandre Chenu, Olivier Serris, Olivier Sigaud et al.

Deep Reinforcement Learning has been successfully applied to learn robotic control. However, the corresponding algorithms struggle when applied to problems where the agent is only rewarded after achieving a complex task. In this context, using demonstrations can significantly speed up the learning process, but demonstrations can be costly to acquire. In this paper, we propose to leverage a sequential bias to learn control policies for complex robotic tasks using a single demonstration. To do so, our method learns a goal-conditioned policy to control a system between successive low-dimensional goals. This sequential goal-reaching approach raises a problem of compatibility between successive goals: we need to ensure that the state resulting from reaching a goal is compatible with the achievement of the following goals. To tackle this problem, we present a new algorithm called DCIL-II. We show that DCIL-II can solve with unprecedented sample efficiency some challenging simulated tasks such as humanoid locomotion and stand-up as well as fast running with a simulated Cassie robot. Our method leveraging sequentiality is a step towards the resolution of complex robotic tasks under minimal specification effort, a key feature for the next generation of autonomous robots.

Alexandre Chenu, Nicolas Perrin-Gilbert, Olivier Sigaud

When cast into the Deep Reinforcement Learning framework, many robotics tasks require solving a long horizon and sparse reward problem, where learning algorithms struggle. In such context, Imitation Learning (IL) can be a powerful approach to bootstrap the learning process. However, most IL methods require several expert demonstrations which can be prohibitively difficult to acquire. Only a handful of IL algorithms have shown efficiency in the context of an extreme low expert data regime where a single expert demonstration is available. In this paper, we present a novel algorithm designed to imitate complex robotic tasks from the states of an expert trajectory. Based on a sequential inductive bias, our method divides the complex task into smaller skills. The skills are learned into a goal-conditioned policy that is able to solve each skill individually and chain skills to solve the entire task. We show that our method imitates a non-holonomic navigation task and scales to a complex simulated robotic manipulation task with very high sample efficiency.

Hugo Caselles-Dupré, Olivier Sigaud, Mohamed Chetouani

We introduce a novel category of GC-agents capable of functioning as both teachers and learners. Leveraging action-based demonstrations and language-based instructions, these agents enhance communication efficiency. We investigate the incorporation of pedagogy and pragmatism, essential elements in human communication and goal achievement, enhancing the agents' teaching and learning capabilities. Furthermore, we explore the impact of combining communication modes (action and language) on learning outcomes, highlighting the benefits of a multi-modal approach.

Zakariae El Asri, Olivier Sigaud, Nicolas Thome

Applying reinforcement learning (RL) to real-world applications requires addressing a trade-off between asymptotic performance, sample efficiency, and inference time. In this work, we demonstrate how to address this triple challenge by leveraging partial physical knowledge about the system dynamics. Our approach involves learning a physics-informed model to boost sample efficiency and generating imaginary trajectories from this model to learn a model-free policy and Q-function. Furthermore, we propose a hybrid planning strategy, combining the learned policy and Q-function with the learned model to enhance time efficiency in planning. Through practical demonstrations, we illustrate that our method improves the compromise between sample efficiency, time efficiency, and performance over state-of-the-art methods.

Theo Cachet, Christopher R. Dance, Olivier Sigaud

Vision-language models (VLMs) have tremendous potential for grounding language, and thus enabling language-conditioned agents (LCAs) to perform diverse tasks specified with text. This has motivated the study of LCAs based on reinforcement learning (RL) with rewards given by rendering images of an environment and evaluating those images with VLMs. If single-task RL is employed, such approaches are limited by the cost and time required to train a policy for each new task. Multi-task RL (MTRL) is a natural alternative, but requires a carefully designed corpus of training tasks and does not always generalize reliably to new tasks. Therefore, this paper introduces a novel decomposition of the problem of building an LCA: first find an environment configuration that has a high VLM score for text describing a task; then use a (pretrained) goal-conditioned policy to reach that configuration. We also explore several enhancements to the speed and quality of VLM-based LCAs, notably, the use of distilled models, and the evaluation of configurations from multiple viewpoints to resolve the ambiguities inherent in a single 2D view. We demonstrate our approach on the Humanoid environment, showing that it results in LCAs that outperform MTRL baselines in zero-shot generalization, without requiring any textual task descriptions or other forms of environment-specific annotation during training. Videos and an interactive demo can be found at https://europe.naverlabs.com/text2control

Ahmed Akakzia, Olivier Sigaud

Although humans live in an open-ended world and endlessly face new challenges, they do not have to learn from scratch each time they face the next one. Rather, they have access to a handful of previously learned skills, which they rapidly adapt to new situations. In artificial intelligence, autotelic agents, which are intrinsically motivated to represent and set their own goals, exhibit promising skill adaptation capabilities. However, these capabilities are highly constrained by their policy and goal space representations. In this paper, we propose to investigate the impact of these representations on the learning and transfer capabilities of autotelic agents. We study different implementations of autotelic agents using four types of Graph Neural Networks policy representations and two types of goal spaces, either geometric or predicate-based. By testing agents on unseen goals, we show that combining object-centered architectures that are expressive enough with semantic relational goals helps learning to reach more difficult goals. We also release our graph-based implementations to encourage further research in this direction.

Mohamed Salim Aissi, Clemence Grislain, Clement Romac et al.

Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.

Zakariae El Asri, Ibrahim Laiche, Clément Rambour et al.

Learning a controller directly on the robot requires extreme sample efficiency. Model-based reinforcement learning (RL) methods are the most sample efficient, but they often suffer from a too long inference time to meet the robot control frequency requirements. In this paper, we address the sample efficiency and inference time challenges with two contributions. First, we define a general framework to deal with inference delays where the slow inference robot controller provides a sequence of actions to feed the control-hungry robotic platform without execution gaps. Then, we compare several RL algorithms in the light of this framework and propose RT-HCP, an algorithm that offers an excellent trade-off between performance, sample efficiency and inference time. We validate the superiority of RT-HCP with experiments where we learn a controller directly on a simple but high frequency FURUTA pendulum platform. Code: github.com/elasriz/RTHCP

Cédric Colas, Olivier Sigaud, Pierre-Yves Oudeyer

Consistently checking the statistical significance of experimental results is one of the mandatory methodological steps to address the so-called "reproducibility crisis" in deep reinforcement learning. In this tutorial paper, we explain how the number of random seeds relates to the probabilities of statistical errors. For both the t-test and the bootstrap confidence interval test, we recall theoretical guidelines to determine the number of random seeds one should use to provide a statistically significant comparison of the performance of two algorithms. Finally, we discuss the influence of deviations from the assumptions usually made by statistical tests. We show that they can lead to inaccurate evaluations of statistical errors and provide guidelines to counter these negative effects. We make our code available to perform the tests.

Cédric Colas, Olivier Sigaud, Pierre-Yves Oudeyer

In continuous action domains, standard deep reinforcement learning algorithms like DDPG suffer from inefficient exploration when facing sparse or deceptive reward problems. Conversely, evolutionary and developmental methods focusing on exploration like Novelty Search, Quality-Diversity or Goal Exploration Processes explore more robustly but are less efficient at fine-tuning policies using gradient descent. In this paper, we present the GEP-PG approach, taking the best of both worlds by sequentially combining a Goal Exploration Process and two variants of DDPG. We study the learning performance of these components and their combination on a low dimensional deceptive reward problem and on the larger Half-Cheetah benchmark. We show that DDPG fails on the former and that GEP-PG improves over the best DDPG variant in both environments. Supplementary videos and discussion can be found at http://frama.link/gep_pg, the code at http://github.com/flowersteam/geppg.

Nazim Bendib, Nicolas Perrin-Gilbert, Olivier Sigaud

Offline zero-shot reinforcement learning (RL) aims to learn agents that optimize unseen reward functions without additional environment interaction. The standard approach to this problem trains task-conditioned policies by sampling task vectors that define linear reward functions over learned state representations. In most existing algorithms, these task vectors are randomly sampled, implicitly assuming this adequately captures the structure of the task space. We argue that doing so leads to suboptimal zero-shot generalization. To address this limitation, we propose extracting task vectors directly from the offline dataset and using them to define the task distribution used for policy training. We introduce a simple and general reward function extraction procedure that integrates into existing offline zero-shot RL algorithms. Across multiple benchmark environments and baselines, our approach improves zero-shot performance by an average of 20%, highlighting the importance of principled task sampling in offline zero-shot RL.

Hamed Rahimi, Clemence Grislain, Adrien Jacquet Cretides et al.

Improving the effectiveness of human-robot interaction requires social robots to accurately infer human goals through robust intention understanding. This challenge is particularly critical in multimodal settings, where agents must integrate heterogeneous signals including text, visual cues to form a coherent interpretation of user intent. This paper presents IntentVLM, a novel two-stage video-language framework designed for open-vocabulary human intention recognition. The approach is inspired by forward-inverse modeling in cognitive science by decomposing intention understanding into goal candidate generation followed by structured inference through selection, effectively reducing hallucinations in latent reasoning. Evaluated on the IntentQA and Inst-IT Bench datasets, IntentVLM achieves state-of-the-art results with up to 80% accuracy, notably surpassing the baseline performance by 30% and matches human performance. Our findings demonstrate that this structured reasoning approach enhances open-vocabulary intention understanding without catastrophic forgetting, offering a robust foundation for human-centered robotics.

Loris Gaven, Clement Romac, Thomas Carta et al.

The past years have seen Large Language Models (LLMs) strive not only as generative models but also as agents solving textual sequential decision-making tasks. When facing complex environments where their zero-shot abilities are insufficient, recent work showed online Reinforcement Learning (RL) could be used for the LLM agent to discover and learn efficient strategies interactively. However, most prior work sticks to on-policy algorithms, which greatly reduces the scope of methods such agents could use for both exploration and exploitation, such as experience replay and hindsight relabeling. Yet, such methods may be key for LLM learning agents, and in particular when designing autonomous intrinsically motivated agents sampling and pursuing their own goals (i.e. autotelic agents). This paper presents and studies an adaptation of Soft Actor-Critic and hindsight relabeling to LLM agents. Our method not only paves the path towards autotelic LLM agents that learn online but can also outperform on-policy methods in more classic multi-goal RL environments.

Loris Gaven, Thomas Carta, Clément Romac et al.

Open-ended learning agents must efficiently prioritize goals in vast possibility spaces, focusing on those that maximize learning progress (LP). When such autotelic exploration is achieved by LLM agents trained with online RL in high-dimensional and evolving goal spaces, a key challenge for LP prediction is modeling one's own competence, a form of metacognitive monitoring. Traditional approaches either require extensive sampling or rely on brittle expert-defined goal groupings. We introduce MAGELLAN, a metacognitive framework that lets LLM agents learn to predict their competence and LP online. By capturing semantic relationships between goals, MAGELLAN enables sample-efficient LP estimation and dynamic adaptation to evolving goal spaces through generalization. In an interactive learning environment, we show that MAGELLAN improves LP prediction efficiency and goal prioritization, being the only method allowing the agent to fully master a large and evolving goal space. These results demonstrate how augmenting LLM agents with a metacognitive ability for LP predictions can effectively scale curriculum learning to open-ended goal spaces.

Mohamed Salim Aissi, Clemence Grislain, Mohamed Chetouani et al.

While Large Language Models (LLMs) excel at reasoning on text and Vision-Language Models (VLMs) are highly effective for visual perception, applying those models for visual instruction-based planning remains a widely open problem. In this paper, we introduce VIPER, a novel framework for multimodal instruction-based planning that integrates VLM-based perception with LLM-based reasoning. Our approach uses a modular pipeline where a frozen VLM generates textual descriptions of image observations, which are then processed by an LLM policy to predict actions based on the task goal. We fine-tune the reasoning module using behavioral cloning and reinforcement learning, improving our agent's decision-making capabilities. Experiments on the ALFWorld benchmark show that VIPER significantly outperforms state-of-the-art visual instruction-based planners while narrowing the gap with purely text-based oracles. By leveraging text as an intermediate representation, VIPER also enhances explainability, paving the way for a fine-grained analysis of perception and reasoning components.

Mohamed Salim Aissi, Clement Romac, Thomas Carta et al.

Reinforcement learning (RL) is a promising approach for aligning large language models (LLMs) knowledge with sequential decision-making tasks. However, few studies have thoroughly investigated the impact on LLM agents capabilities of fine-tuning them with RL in a specific environment. In this paper, we propose a novel framework to analyze the sensitivity of LLMs to prompt formulations following RL training in a textual environment. Our findings reveal that the performance of LLMs degrades when faced with prompt formulations different from those used during the RL training phase. Besides, we analyze the source of this sensitivity by examining the model's internal representations and salient tokens. Finally, we propose to use a contrastive loss to mitigate this sensitivity and improve the robustness and generalization capabilities of LLMs.

Thomas Carta, Clément Romac, Loris Gaven et al.

Open-ended AI agents need to be able to learn efficiently goals of increasing complexity, abstraction and heterogeneity over their lifetime. Beyond sampling efficiently their own goals, autotelic agents specifically need to be able to keep the growing complexity of goals under control, limiting the associated growth in sample and computational complexity. To adress this challenge, recent approaches have leveraged hierarchical reinforcement learning (HRL) and language, capitalizing on its compositional and combinatorial generalization capabilities to acquire temporally extended reusable behaviours. Existing approaches use expert defined spaces of subgoals over which they instantiate a hierarchy, and often assume pre-trained associated low-level policies. Such designs are inadequate in open-ended scenarios, where goal spaces naturally diversify across a broad spectrum of difficulties. We introduce HERAKLES, a framework that enables a two-level hierarchical autotelic agent to continuously compile mastered goals into the low-level policy, executed by a small, fast neural network, dynamically expanding the set of subgoals available to the high-level policy. We train a Large Language Model (LLM) to serve as the high-level controller, exploiting its strengths in goal decomposition and generalization to operate effectively over this evolving subgoal space. We evaluate HERAKLES in the open-ended Crafter environment and show that it scales effectively with goal complexity, improves sample efficiency through skill compilation, and enables the agent to adapt robustly to novel challenges over time.

Nicolas Castanet, Olivier Sigaud, Sylvain Lamprier

Goal-Conditioned Reinforcement Learning (GCRL) enables agents to autonomously acquire diverse behaviors, but faces major challenges in visual environments due to high-dimensional, semantically sparse observations. In the online setting, where agents learn representations while exploring, the latent space evolves with the agent's policy, to capture newly discovered areas of the environment. However, without incentivization to maximize state coverage in the representation, classical approaches based on auto-encoders may converge to latent spaces that over-represent a restricted set of states frequently visited by the agent. This is exacerbated in an intrinsic motivation setting, where the agent uses the distribution encoded in the latent space to sample the goals it learns to master. To address this issue, we propose to progressively enforce distributional shifts towards a uniform distribution over the full state space, to ensure a full coverage of skills that can be learned in the environment. We introduce DRAG (Distributionally Robust Auto-Encoding for GCRL), a method that combines the $β$-VAE framework with Distributionally Robust Optimization. DRAG leverages an adversarial neural weighter of training states of the VAE, to account for the mismatch between the current data distribution and unseen parts of the environment. This allows the agent to construct semantically meaningful latent spaces beyond its immediate experience. Our approach improves state space coverage and downstream control performance on hard exploration environments such as mazes and robotic control involving walls to bypass, without pre-training nor prior environment knowledge.

Olivier Serris, Stéphane Doncieux, Olivier Sigaud

Several hierarchical reinforcement learning methods leverage planning to create a graph or sequences of intermediate goals, guiding a lower-level goal-conditioned (GC) policy to reach some final goals. The low-level policy is typically conditioned on the current goal, with the aim of reaching it as quickly as possible. However, this approach can fail when an intermediate goal can be reached in multiple ways, some of which may make it impossible to continue toward subsequent goals. To address this issue, we introduce two instances of Markov Decision Process (MDP) where the optimization objective favors policies that not only reach the current goal but also subsequent ones. In the first, the agent is conditioned on both the current and final goals, while in the second, it is conditioned on the next two goals in the sequence. We conduct a series of experiments on navigation and pole-balancing tasks in which sequences of intermediate goals are given. By evaluating policies trained with TD3+HER on both the standard GC-MDP and our proposed MDPs, we show that, in most cases, conditioning on the next two goals improves stability and sample efficiency over other approaches.

Alexandre Chenu, Olivier Serris, Olivier Sigaud et al.

Demonstrations are commonly used to speed up the learning process of Deep Reinforcement Learning algorithms. To cope with the difficulty of accessing multiple demonstrations, some algorithms have been developed to learn from a single demonstration. In particular, the Divide & Conquer Imitation Learning algorithms leverage a sequential bias to learn a control policy for complex robotic tasks using a single state-based demonstration. The latest version, DCIL-II demonstrates remarkable sample efficiency. This novel method operates within an extended Goal-Conditioned Reinforcement Learning framework, ensuring compatibility between intermediate and subsequent goals extracted from the demonstration. However, a fundamental limitation arises from the assumption that the system can be reset to specific states along the demonstrated trajectory, confining the application to simulated systems. In response, we introduce an extension called Single-Reset DCIL (SR-DCIL), designed to overcome this constraint by relying on a single initial state reset rather than sequential resets. To address this more challenging setting, we integrate two mechanisms inspired by the Learning from Demonstrations literature, including a Demo-Buffer and Value Cloning, to guide the agent toward compatible success states. In addition, we introduce Approximate Goal Switching to facilitate training to reach goals distant from the reset state. Our paper makes several contributions, highlighting the importance of the reset assumption in DCIL-II, presenting the mechanisms of SR-DCIL variants and evaluating their performance in challenging robotic tasks compared to DCIL-II. In summary, this work offers insights into the significance of reset assumptions in the framework of DCIL and proposes SR-DCIL, a first step toward a versatile algorithm capable of learning control policies under a weaker reset assumption.

Hugo Caselles-Dupré, Mohamed Chetouani, Olivier Sigaud

When demonstrating a task, human tutors pedagogically modify their behavior by either "showing" the task rather than just "doing" it (exaggerating on relevant parts of the demonstration) or by giving demonstrations that best disambiguate the communicated goal. Analogously, human learners pragmatically infer the communicative intent of the tutor: they interpret what the tutor is trying to teach them and deduce relevant information for learning. Without such mechanisms, traditional Learning from Demonstration (LfD) algorithms will consider such demonstrations as sub-optimal. In this paper, we investigate the implementation of such mechanisms in a tutor-learner setup where both participants are artificial agents in an environment with multiple goals. Using pedagogy from the tutor and pragmatism from the learner, we show substantial improvements over standard learning from demonstrations.

Ahmed Akakzia, Olivier Serris, Olivier Sigaud et al.

In the quest for autonomous agents learning open-ended repertoires of skills, most works take a Piagetian perspective: learning trajectories are the results of interactions between developmental agents and their physical environment. The Vygotskian perspective, on the other hand, emphasizes the centrality of the socio-cultural environment: higher cognitive functions emerge from transmissions of socio-cultural processes internalized by the agent. This paper argues that both perspectives could be coupled within the learning of autotelic agents to foster their skill acquisition. To this end, we make two contributions: 1) a novel social interaction protocol called Help Me Explore (HME), where autotelic agents can benefit from both individual and socially guided exploration. In social episodes, a social partner suggests goals at the frontier of the learning agent knowledge. In autotelic episodes, agents can either learn to master their own discovered goals or autonomously rehearse failed social goals; 2) GANGSTR, a graph-based autotelic agent for manipulation domains capable of decomposing goals into sequences of intermediate sub-goals. We show that when learning within HME, GANGSTR overcomes its individual learning limits by mastering the most complex configurations (e.g. stacks of 5 blocks) with only few social interventions.

Olivier Sigaud, Ahmed Akakzia, Hugo Caselles-Dupré et al.

Autonomous discovery and direct instruction are two distinct sources of learning in children but education sciences demonstrate that mixed approaches such as assisted discovery or guided play result in improved skill acquisition. In the field of Artificial Intelligence, these extremes respectively map to autonomous agents learning from their own signals and interactive learning agents fully taught by their teachers. In between should stand teachable autotelic agents (TAA): agents that learn from both internal and teaching signals to benefit from the higher efficiency of assisted discovery. Designing such agents will enable real-world non-expert users to orient the learning trajectories of agents towards their expectations. More fundamentally, this may also be a key step to build agents with human-level intelligence. This paper presents a roadmap towards the design of teachable autonomous agents. Building on developmental psychology and education sciences, we start by identifying key features enabling assisted discovery processes in child-tutor interactions. This leads to the production of a checklist of features that future TAA will need to demonstrate. The checklist allows us to precisely pinpoint the various limitations of current reinforcement learning agents and to identify the promising first steps towards TAA. It also shows the way forward by highlighting key research directions towards the design or autonomous agents that can be taught by ordinary people via natural pedagogy.

Alexandre Chenu, Nicolas Perrin-Gilbert, Stéphane Doncieux et al.

Reinforcement learning agents need a reward signal to learn successful policies. When this signal is sparse or the corresponding gradient is deceptive, such agents need a dedicated mechanism to efficiently explore their search space without relying on the reward. Looking for a large diversity of behaviors or using Motion Planning (MP) algorithms are two options in this context. In this paper, we build on the common roots between these two options to investigate the properties of two diversity search algorithms, the Novelty Search and the Goal Exploration Process algorithms. These algorithms look for diversity in an outcome space or behavioral space which is generally hand-designed to represent what matters for a given task. The relation to MP algorithms reveals that the smoothness, or lack of smoothness of the mapping between the policy parameter space and the outcome space plays a key role in the search efficiency. In particular, we show empirically that, if the mapping is smooth enough, i.e. if two close policies in the parameter space lead to similar outcomes, then diversity algorithms tend to inherit exploration properties of MP algorithms. By contrast, if it is not, diversity algorithms lose these properties and their performance strongly depends on specific heuristics, notably filtering mechanisms that discard some of the explored policies.

Cédric Colas, Tristan Karch, Olivier Sigaud et al.

Building autonomous machines that can explore open-ended environments, discover possible interactions and build repertoires of skills is a general objective of artificial intelligence. Developmental approaches argue that this can only be achieved by $autotelic$ $agents$: intrinsically motivated learning agents that can learn to represent, generate, select and solve their own problems. In recent years, the convergence of developmental approaches with deep reinforcement learning (RL) methods has been leading to the emergence of a new field: $developmental$ $reinforcement$ $learning$. Developmental RL is concerned with the use of deep RL algorithms to tackle a developmental problem -- the $intrinsically$ $motivated$ $acquisition$ $of$ $open$-$ended$ $repertoires$ $of$ $skills$. The self-generation of goals requires the learning of compact goal encodings as well as their associated goal-achievement functions. This raises new challenges compared to standard RL algorithms originally designed to tackle pre-defined sets of goals using external reward signals. The present paper introduces developmental RL and proposes a computational framework based on goal-conditioned RL to tackle the intrinsically motivated skills acquisition problem. It proceeds to present a typology of the various goal representations used in the literature, before reviewing existing methods to learn to represent and prioritize goals in autonomous systems. We finally close the paper by discussing some open challenges in the quest of intrinsically motivated skills acquisition.

Louis Monier, Jakub Kmec, Alexandre Laterre et al.

Offline Reinforcement Learning (RL) aims to turn large datasets into powerful decision-making engines without any online interactions with the environment. This great promise has motivated a large amount of research that hopes to replicate the success RL has experienced in simulation settings. This work ambitions to reflect upon these efforts from a practitioner viewpoint. We start by discussing the dataset properties that we hypothesise can characterise the type of offline methods that will be the most successful. We then verify these claims through a set of experiments and designed datasets generated from environments with both discrete and continuous action spaces. We experimentally validate that diversity and high-return examples in the data are crucial to the success of offline RL and show that behavioural cloning remains a strong contender compared to its contemporaries. Overall, this work stands as a tutorial to help people build their intuition on today's offline RL methods and their applicability.

Thomas Pierrot, Nicolas Perrin, Feryal Behbahani et al.

We propose a novel solution to challenging sparse-reward, continuous control problems that require hierarchical planning at multiple levels of abstraction. Our solution, dubbed AlphaNPI-X, involves three separate stages of learning. First, we use off-policy reinforcement learning algorithms with experience replay to learn a set of atomic goal-conditioned policies, which can be easily repurposed for many tasks. Second, we learn self-models describing the effect of the atomic policies on the environment. Third, the self-models are harnessed to learn recursive compositional programs with multiple levels of abstraction. The key insight is that the self-models enable planning by imagination, obviating the need for interaction with the world when learning higher-level compositional programs. To accomplish the third stage of learning, we extend the AlphaNPI algorithm, which applies AlphaZero to learn recursive neural programmer-interpreters. We empirically show that AlphaNPI-X can effectively learn to tackle challenging sparse manipulation tasks, such as stacking multiple blocks, where powerful model-free baselines fail.

Thomas Pierrot, Valentin Macé, Félix Chalumeau et al.

A fascinating aspect of nature lies in its ability to produce a large and diverse collection of organisms that are all high-performing in their niche. By contrast, most AI algorithms focus on finding a single efficient solution to a given problem. Aiming for diversity in addition to performance is a convenient way to deal with the exploration-exploitation trade-off that plays a central role in learning. It also allows for increased robustness when the returned collection contains several working solutions to the considered problem, making it well-suited for real applications such as robotics. Quality-Diversity (QD) methods are evolutionary algorithms designed for this purpose. This paper proposes a novel algorithm, QDPG, which combines the strength of Policy Gradient algorithms and Quality Diversity approaches to produce a collection of diverse and high-performing neural policies in continuous control environments. The main contribution of this work is the introduction of a Diversity Policy Gradient (DPG) that exploits information at the time-step level to drive policies towards more diversity in a sample-efficient manner. Specifically, QDPG selects neural controllers from a MAP-Elites grid and uses two gradient-based mutation operators to improve both quality and diversity. Our results demonstrate that QDPG is significantly more sample-efficient than its evolutionary competitors.

Ahmed Akakzia, Cédric Colas, Pierre-Yves Oudeyer et al.

We are interested in the autonomous acquisition of repertoires of skills. Language-conditioned reinforcement learning (LC-RL) approaches are great tools in this quest, as they allow to express abstract goals as sets of constraints on the states. However, most LC-RL agents are not autonomous and cannot learn without external instructions and feedback. Besides, their direct language condition cannot account for the goal-directed behavior of pre-verbal infants and strongly limits the expression of behavioral diversity for a given language input. To resolve these issues, we propose a new conceptual approach to language-conditioned RL: the Language-Goal-Behavior architecture (LGB). LGB decouples skill learning and language grounding via an intermediate semantic representation of the world. To showcase the properties of LGB, we present a specific implementation called DECSTR. DECSTR is an intrinsically motivated learning agent endowed with an innate semantic representation describing spatial relations between physical objects. In a first stage (G -> B), it freely explores its environment and targets self-generated semantic configurations. In a second stage (L -> G), it trains a language-conditioned goal generator to generate semantic goals that match the constraints expressed in language-based inputs. We showcase the additional properties of LGB w.r.t. both an end-to-end LC-RL approach and a similar approach leveraging non-semantic, continuous intermediate representations. Intermediate semantic representations help satisfy language commands in a diversity of ways, enable strategy switching after a failure and facilitate language grounding.

Cédric Colas, Ahmed Akakzia, Pierre-Yves Oudeyer et al.

In the real world, linguistic agents are also embodied agents: they perceive and act in the physical world. The notion of Language Grounding questions the interactions between language and embodiment: how do learning agents connect or ground linguistic representations to the physical world ? This question has recently been approached by the Reinforcement Learning community under the framework of instruction-following agents. In these agents, behavioral policies or reward functions are conditioned on the embedding of an instruction expressed in natural language. This paper proposes another approach: using language to condition goal generators. Given any goal-conditioned policy, one could train a language-conditioned goal generator to generate language-agnostic goals for the agent. This method allows to decouple sensorimotor learning from language acquisition and enable agents to demonstrate a diversity of behaviors for any given instruction. We propose a particular instantiation of this approach and demonstrate its benefits.

Stephane Doncieux, Nicolas Bredeche, Léni Le Goff et al.

Robots are still limited to controlled conditions, that the robot designer knows with enough details to endow the robot with the appropriate models or behaviors. Learning algorithms add some flexibility with the ability to discover the appropriate behavior given either some demonstrations or a reward to guide its exploration with a reinforcement learning algorithm. Reinforcement learning algorithms rely on the definition of state and action spaces that define reachable behaviors. Their adaptation capability critically depends on the representations of these spaces: small and discrete spaces result in fast learning while large and continuous spaces are challenging and either require a long training period or prevent the robot from converging to an appropriate behavior. Beside the operational cycle of policy execution and the learning cycle, which works at a slower time scale to acquire new policies, we introduce the redescription cycle, a third cycle working at an even slower time scale to generate or adapt the required representations to the robot, its environment and the task. We introduce the challenges raised by this cycle and we present DREAM (Deferred Restructuring of Experience in Autonomous Machines), a developmental cognitive architecture to bootstrap this redescription process stage by stage, build new state representations with appropriate motivations, and transfer the acquired knowledge across domains or tasks or even across robots. We describe results obtained so far with this approach and end up with a discussion of the questions it raises in Neuroscience.

Guillaume Matheron, Nicolas Perrin, Olivier Sigaud

The exploration-exploitation trade-off is at the heart of reinforcement learning (RL). However, most continuous control benchmarks used in recent RL research only require local exploration. This led to the development of algorithms that have basic exploration capabilities, and behave poorly in benchmarks that require more versatile exploration. For instance, as demonstrated in our empirical study, state-of-the-art RL algorithms such as DDPG and TD3 are unable to steer a point mass in even small 2D mazes. In this paper, we propose a new algorithm called "Plan, Backplay, Chain Skills" (PBCS) that combines motion planning and reinforcement learning to solve hard exploration environments. In a first phase, a motion planning algorithm is used to find a single good trajectory, then an RL algorithm is trained using a curriculum derived from the trajectory, by combining a variant of the Backplay algorithm and skill chaining. We show that this method outperforms state-of-the-art RL algorithms in 2D maze environments of various sizes, and is able to improve on the trajectory obtained by the motion planning phase.

Aloïs Pourchot, Alexis Ducarouge, Olivier Sigaud

Weight-sharing (WS) has recently emerged as a paradigm to accelerate the automated search for efficient neural architectures, a process dubbed Neural Architecture Search (NAS). Although very appealing, this framework is not without drawbacks and several works have started to question its capabilities on small hand-crafted benchmarks. In this paper, we take advantage of the \nasbench dataset to challenge the efficiency of WS on a representative search space. By comparing a SOTA WS approach to a plain random search we show that, despite decent correlations between evaluations using weight-sharing and standalone ones, WS is only rarely significantly helpful to NAS. In particular we highlight the impact of the search space itself on the benefits.

Guillaume Matheron, Nicolas Perrin, Olivier Sigaud

In environments with continuous state and action spaces, state-of-the-art actor-critic reinforcement learning algorithms can solve very complex problems, yet can also fail in environments that seem trivial, but the reason for such failures is still poorly understood. In this paper, we contribute a formal explanation of these failures in the particular case of sparse reward and deterministic environments. First, using a very elementary control problem, we illustrate that the learning process can get stuck into a fixed point corresponding to a poor solution. Then, generalizing from the studied example, we provide a detailed analysis of the underlying mechanisms which results in a new understanding of one of the convergence regimes of these algorithms. The resulting perspective casts a new light on already existing solutions to the issues we have highlighted, and suggests other potential approaches.

Thomas Pierrot, Guillaume Ligner, Scott Reed et al.

We propose a novel reinforcement learning algorithm, AlphaNPI, that incorporates the strengths of Neural Programmer-Interpreters (NPI) and AlphaZero. NPI contributes structural biases in the form of modularity, hierarchy and recursion, which are helpful to reduce sample complexity, improve generalization and increase interpretability. AlphaZero contributes powerful neural network guided search algorithms, which we augment with recursion. AlphaNPI only assumes a hierarchical program specification with sparse rewards: 1 when the program execution satisfies the specification, and 0 otherwise. Using this specification, AlphaNPI is able to train NPI models effectively with RL for the first time, completely eliminating the need for strong supervision in the form of execution traces. The experiments show that AlphaNPI can sort as well as previous strongly supervised NPI variants. The AlphaNPI agent is also trained on a Tower of Hanoi puzzle with two disks and is shown to generalize to puzzles with an arbitrary number of disk

Cédric Colas, Olivier Sigaud, Pierre-Yves Oudeyer

Consistently checking the statistical significance of experimental results is the first mandatory step towards reproducible science. This paper presents a hitchhiker's guide to rigorous comparisons of reinforcement learning algorithms. After introducing the concepts of statistical testing, we review the relevant statistical tests and compare them empirically in terms of false positive rate and statistical power as a function of the sample size (number of seeds) and effect size. We further investigate the robustness of these tests to violations of the most common hypotheses (normal distributions, same distributions, equal variances). Beside simulations, we compare empirical distributions obtained by running Soft-Actor Critic and Twin-Delayed Deep Deterministic Policy Gradient on Half-Cheetah. We conclude by providing guidelines and code to perform rigorous comparisons of RL algorithm performances.

Chenyang Zhao, Olivier Sigaud, Freek Stulp et al.

Deep Reinforcement Learning has shown great success in a variety of control tasks. However, it is unclear how close we are to the vision of putting Deep RL into practice to solve real world problems. In particular, common practice in the field is to train policies on largely deterministic simulators and to evaluate algorithms through training performance alone, without a train/test distinction to ensure models generalise and are not overfitted. Moreover, it is not standard practice to check for generalisation under domain shift, although robustness to such system change between training and testing would be necessary for real-world Deep RL control, for example, in robotics. In this paper we study these issues by first characterising the sources of uncertainty that provide generalisation challenges in Deep RL. We then provide a new benchmark and thorough empirical evaluation of generalisation challenges for state of the art Deep RL methods. In particular, we show that, if generalisation is the goal, then common practice of evaluating algorithms based on their training performance leads to the wrong conclusions about algorithm choice. Finally, we evaluate several techniques for improving generalisation and draw conclusions about the most robust techniques to date.

Anis Najar, Olivier Sigaud, Mohamed Chetouani

In this paper, we propose a framework that enables a human teacher to shape a robot behaviour by interactively providing it with unlabeled instructions. We ground the meaning of instruction signals in the task-learning process, and use them simultaneously for guiding the latter. We implement our framework as a modular architecture, named TICS (Task-Instruction-Contingency-Shaping) that combines different information sources: a predefined reward function, human evaluative feedback and unlabeled instructions. This approach provides a novel perspective for robotic task learning that lies between Reinforcement Learning and Supervised Learning paradigms. We evaluate our framework both in simulation and with a real robot. The experimental results demonstrate the effectiveness of our framework in accelerating the task-learning process and in reducing the number of required teaching signals.

Pierre Fournier, Olivier Sigaud, Cédric Colas et al.

In this paper we study a new reinforcement learning setting where the environment is non-rewarding, contains several possibly related objects of various controllability, and where an apt agent Bob acts independently, with non-observable intentions. We argue that this setting defines a realistic scenario and we present a generic discrete-state discrete-action model of such environments. To learn in this environment, we propose an unsupervised reinforcement learning agent called CLIC for Curriculum Learning and Imitation for Control. CLIC learns to control individual objects in its environment, and imitates Bob's interactions with these objects. It selects objects to focus on when training and imitating by maximizing its learning progress. We show that CLIC is an effective baseline in our new setting. It can effectively observe Bob to gain control of objects faster, even if Bob is not explicitly teaching. It can also follow Bob when he acts as a mentor and provides ordered demonstrations. Finally, when Bob controls objects that the agent cannot, or in presence of a hierarchy between objects in the environment, we show that CLIC ignores non-reproducible and already mastered interactions with objects, resulting in a greater benefit from imitation.

Thomas Pierrot, Nicolas Perrin, Olivier Sigaud

In this paper, we provide an overview of first-order and second-order variants of the gradient descent method that are commonly used in machine learning. We propose a general framework in which 6 of these variants can be interpreted as different instances of the same approach. They are the vanilla gradient descent, the classical and generalized Gauss-Newton methods, the natural gradient descent method, the gradient covariance matrix approach, and Newton's method. Besides interpreting these methods within a single framework, we explain their specificities and show under which conditions some of them coincide.

Cédric Colas, Pierre Fournier, Olivier Sigaud et al.

In open-ended environments, autonomous learning agents must set their own goals and build their own curriculum through an intrinsically motivated exploration. They may consider a large diversity of goals, aiming to discover what is controllable in their environments, and what is not. Because some goals might prove easy and some impossible, agents must actively select which goal to practice at any moment, to maximize their overall mastery on the set of learnable goals. This paper proposes CURIOUS, an algorithm that leverages 1) a modular Universal Value Function Approximator with hindsight learning to achieve a diversity of goals of different kinds within a unique policy and 2) an automated curriculum learning mechanism that biases the attention of the agent towards goals maximizing the absolute learning progress. Agents focus sequentially on goals of increasing complexity, and focus back on goals that are being forgotten. Experiments conducted in a new modular-goal robotic environment show the resulting developmental self-organization of a learning curriculum, and demonstrate properties of robustness to distracting goals, forgetting and changes in body properties.