Successfully addressing a wide variety of tasks is a core ability of autonomous agents, which requires flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the underlying perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, in this work, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pre-trained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the policy and visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks of the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given visual demonstrations.

Learning robot navigation strategies among pedestrian is crucial for domain based applications. Combining perception, planning and prediction allows us to model the interactions between robots and pedestrians, resulting in impressive outcomes especially with recent approaches based on deep reinforcement learning (RL). However, these works do not consider multi-robot scenarios. In this paper, we present MultiSoc, a new method for learning multi-agent socially aware navigation strategies using RL. Inspired by recent works on multi-agent deep RL, our method leverages graph-based representation of agent interactions, combining the positions and fields of view of entities (pedestrians and agents). Each agent uses a model based on two Graph Neural Network combined with attention mechanisms. First an edge-selector produces a sparse graph, then a crowd coordinator applies node attention to produce a graph representing the influence of each entity on the others. This is incorporated into a model-free RL framework to learn multi-agent policies. We evaluate our approach on simulation and provide a series of experiments in a set of various conditions (number of agents / pedestrians). Empirical results show that our method learns faster than social navigation deep RL mono-agent techniques, and enables efficient multi-agent implicit coordination in challenging crowd navigation with multiple heterogeneous humans. Furthermore, by incorporating customizable meta-parameters, we can adjust the neighborhood density to take into account in our navigation strategy.

Implicit representations such as Neural Radiance Fields (NeRF) have been shown to be very effective at novel view synthesis. However, these models typically require manual and careful human data collection for training. In this paper, we present AutoNeRF, a method to collect data required to train NeRFs using autonomous embodied agents. Our method allows an agent to explore an unseen environment efficiently and use the experience to build an implicit map representation autonomously. We compare the impact of different exploration strategies including handcrafted frontier-based exploration, end-to-end and modular approaches composed of trained high-level planners and classical low-level path followers. We train these models with different reward functions tailored to this problem and evaluate the quality of the learned representations on four different downstream tasks: classical viewpoint rendering, map reconstruction, planning, and pose refinement. Empirical results show that NeRFs can be trained on actively collected data using just a single episode of experience in an unseen environment, and can be used for several downstream robotic tasks, and that modular trained exploration models outperform other classical and end-to-end baselines. Finally, we show that AutoNeRF can reconstruct large-scale scenes, and is thus a useful tool to perform scene-specific adaptation as the produced 3D environment models can be loaded into a simulator to fine-tune a policy of interest.

Understanding and mapping a new environment are core abilities of any autonomously navigating agent. While classical robotics usually estimates maps in a stand-alone manner with SLAM variants, which maintain a topological or metric representation, end-to-end learning of navigation keeps some form of memory in a neural network. Networks are typically imbued with inductive biases, which can range from vectorial representations to birds-eye metric tensors or topological structures. In this work, we propose to structure neural networks with two neural implicit representations, which are learned dynamically during each episode and map the content of the scene: (i) the Semantic Finder predicts the position of a previously seen queried object; (ii) the Occupancy and Exploration Implicit Representation encapsulates information about explored area and obstacles, and is queried with a novel global read mechanism which directly maps from function space to a usable embedding space. Both representations are leveraged by an agent trained with Reinforcement Learning (RL) and learned online during each episode. We evaluate the agent on Multi-Object Navigation and show the high impact of using neural implicit representations as a memory source.

In the context of visual navigation, the capacity to map a novel environment is necessary for an agent to exploit its observation history in the considered place and efficiently reach known goals. This ability can be associated with spatial reasoning, where an agent is able to perceive spatial relationships and regularities, and discover object characteristics. Recent work introduces learnable policies parametrized by deep neural networks and trained with Reinforcement Learning (RL). In classical RL setups, the capacity to map and reason spatially is learned end-to-end, from reward alone. In this setting, we introduce supplementary supervision in the form of auxiliary tasks designed to favor the emergence of spatial perception capabilities in agents trained for a goal-reaching downstream objective. We show that learning to estimate metrics quantifying the spatial relationships between an agent at a given location and a goal to reach has a high positive impact in Multi-Object Navigation settings. Our method significantly improves the performance of different baseline agents, that either build an explicit or implicit representation of the environment, even matching the performance of incomparable oracle agents taking ground-truth maps as input. A learning-based agent from the literature trained with the proposed auxiliary losses was the winning entry to the Multi-Object Navigation Challenge, part of the CVPR 2021 Embodied AI Workshop.

Traditionally, an agent maximizes a reward defined according to the task to perform: it may be a score when the agent learns to solve a game or a distance function when the agent learns to reach a goal. The reward is then considered as extrinsic (or as a feedback) because the reward function is provided expertly and specifically for the task. With an extrinsic reward, many spectacular results have been obtained on Atari game [Bellemare et al. 2015] with the Deep Q-network (DQN) [Mnih et al. 2015] through the integration of deep learning to RL, leading to deep reinforcement learning (DRL). However, despite the recent improvements of DRL approaches, they turn out to be most of the time unsuccessful when the rewards are scattered in the environment, as the agent is then unable to learn the desired behavior for the targeted task [Francois-Lavet et al. 2018]. Moreover, the behaviors learned by the agent are hardly reusable, both within the same task and across many different tasks [Francois-Lavet et al. 2018]. It is difficult for an agent to generalize the learnt skills to make high-level decisions in the environment. For example, such skill could be go to the door using primitive actions consisting in moving in the four cardinal directions; or even to move forward controlling different joints of a humanoid robot like in the robotic simulator MuJoCo [Todorov et al. 2012]. On another side, unlike RL, developmental learning [Cangelosi and Schlesinger 2018; Oudeyer and Smith 2016; Piaget and Cook 1952] is based on the trend that babies, or more broadly organisms, acquire new skill while spontaneously exploring their environment [Barto 2013; Gopnik et al. 1999].

Taking inspiration from developmental learning, we present a novel reinforcement learning architecture which hierarchically learns and represents self-generated skills in an end-to-end way. With this architecture, an agent focuses only on task-rewarded skills while keeping the learning process of skills bottom-up. This bottom-up approach allows to learn skills that 1- are transferable across tasks, 2- improves exploration when rewards are sparse. To do so, we combine a previously defined mutual information objective with a novel curriculum learning algorithm, creating an unlimited and explorable tree of skills. We test our agent on simple gridworld environments to understand and visualize how the agent distinguishes between its skills. Then we show that our approach can scale on more difficult MuJoCo environments in which our agent is able to build a representation of skills which improve over a baseline both transfer learning and exploration when rewards are sparse.

Despite numerous research work in reinforcement learning (RL) and the recent successes obtained by combining it with deep learning, deep reinforcement learning (DRL) is still facing many challenges. Some of them, like the ability to abstract actions or the difficulty to explore the environment with sparse rewards, can be addressed by the use of intrinsic motivation. In this article, we provide a survey on the role of intrinsic motivation in DRL. We categorize the different kinds of intrinsic motivations and detail their interests and limitations. Our investigation shows that the combination of DRL and intrinsic motivation enables to learn more complicated and more generalisable behaviours than standard DRL. We provide an in-depth analysis describing learning modules through an unifying scheme composed of information theory, compression theory and reinforcement learning. We then explain how these modules could serve as building blocks over a complete developmental architecture, highlighting the numerous outlooks of the domain.

Recent works on multi-agent sequential decision making using decentralized partially observable Markov decision processes have been concerned with interaction-oriented resolution techniques and provide promising results. These techniques take advantage of local interactions and coordination. In this paper, we propose an approach based on an interaction-oriented resolution of decentralized decision makers. To this end, distributed value functions (DVF) have been used by decoupling the multi-agent problem into a set of individual agent problems. However existing DVF techniques assume permanent and free communication between the agents. In this paper, we extend the DVF methodology to address full local observability, limited share of information and communication breaks. We apply our new DVF in a real-world application consisting of multi-robot exploration where each robot computes locally a strategy that minimizes the interactions between the robots and maximizes the space coverage of the team even under communication constraints. Our technique has been implemented and evaluated in simulation and in real-world scenarios during a robotic challenge for the exploration and mapping of an unknown environment. Experimental results from real-world scenarios and from the challenge are given where our system was vice-champion.

We are motivated by building a system for an autonomous robot companion that collaborates with a human partner for achieving a common mission. The objective of the robot is to infer the human's preferences upon the tasks of the mission so as to collaborate with the human by achieving human's non-favorite tasks. Inspired by recent researches about the recognition of human's intention, we propose a unified model that allows the robot to switch accurately between verbal and non-verbal interactions. Our system unifies an epistemic partially observable Markov decision process (POMDP) that is a human-robot spoken dialog system aiming at disambiguating the human's preferences and an intuitive human-robot collaboration consisting in inferring human's intention based on the observed human actions. The beliefs over human's preferences computed during the dialog are then reinforced in the course of the task execution by the intuitive interaction. Our unified model helps the robot inferring the human's preferences and deciding which tasks to perform to effectively satisfy these preferences. The robot is also able to adjust its plan rapidly in case of sudden changes in the human's preferences and to switch between both kind of interactions. Experimental results on a scenario inspired from robocup@home outline various specific behaviors of the robot during the collaborative mission.