We discuss scenarios and branch points to four major possible consequences regarding future machine intelligence; 1) the singleton scenario where the first and only super-intelligence acquires a decisive strategic advantage, 2) the multipolar scenario where the singleton scenario is not technically denied but political or other factors in human society or multi-agent interactions between the intelligent agents prevent a single agent from gaining a decisive strategic advantage, 3) the ecosystem scenario where the singleton scenario is denied and many autonomous intelligent agents operate in such a way that they are interdependent and virtually unstoppable, and 4) the upper-bound scenario where cognitive capabilities that can be achieved by human-designed intelligent agents or their descendants are inherently limited to the sub-human level. We identify six major constraints that can form branch points to these scenarios; (1) constraints on autonomy, (2) constraints on the ability to improve self-structure, (3) constraints related to thermodynamic efficiency, (4) constraints on updating physical infrastructure, (5) constraints on relative advantage, and (6) constraints on locality.

Multi-agent reinforcement learning typically suffers from the problem of sample inefficiency, where learning suitable policies involves the use of many data samples. Learning from external demonstrators is a possible solution that mitigates this problem. However, most prior approaches in this area assume the presence of a single demonstrator. Leveraging multiple knowledge sources (i.e., advisors) with expertise in distinct aspects of the environment could substantially speed up learning in complex environments. This paper considers the problem of simultaneously learning from multiple independent advisors in multi-agent reinforcement learning. The approach leverages a two-level Q-learning architecture, and extends this framework from single-agent to multi-agent settings. We provide principled algorithms that incorporate a set of advisors by both evaluating the advisors at each state and subsequently using the advisors to guide action selection. We also provide theoretical convergence and sample complexity guarantees. Experimentally, we validate our approach in three different test-beds and show that our algorithms give better performances than baselines, can effectively integrate the combined expertise of different advisors, and learn to ignore bad advice.

A unique challenge in Multi-Agent Reinforcement Learning (MARL) is the curse of multiagency, where the description length of the game as well as the complexity of many existing learning algorithms scale exponentially with the number of agents. While recent works successfully address this challenge under the model of tabular Markov Games, their mechanisms critically rely on the number of states being finite and small, and do not extend to practical scenarios with enormous state spaces where function approximation must be used to approximate value functions or policies. This paper presents the first line of MARL algorithms that provably resolve the curse of multiagency under function approximation. We design a new decentralized algorithm -- V-Learning with Policy Replay, which gives the first polynomial sample complexity results for learning approximate Coarse Correlated Equilibria (CCEs) of Markov Games under decentralized linear function approximation. Our algorithm always outputs Markov CCEs, and achieves an optimal rate of $\widetilde{\mathcal{O}}(\epsilon^{-2})$ for finding $\epsilon$-optimal solutions. Also, when restricted to the tabular case, our result improves over the current best decentralized result $\widetilde{\mathcal{O}}(\epsilon^{-3})$ for finding Markov CCEs. We further present an alternative algorithm -- Decentralized Optimistic Policy Mirror Descent, which finds policy-class-restricted CCEs using a polynomial number of samples. In exchange for learning a weaker version of CCEs, this algorithm applies to a wider range of problems under generic function approximation, such as linear quadratic games and MARL problems with low ''marginal'' Eluder dimension.

We aim to understand how people assess human likeness in navigation produced by people and artificially intelligent (AI) agents in a video game. To this end, we propose a novel AI agent with the goal of generating more human-like behavior. We collect hundreds of crowd-sourced assessments comparing the human-likeness of navigation behavior generated by our agent and baseline AI agents with human-generated behavior. Our proposed agent passes a Turing Test, while the baseline agents do not. By passing a Turing Test, we mean that human judges could not quantitatively distinguish between videos of a person and an AI agent navigating. To understand what people believe constitutes human-like navigation, we extensively analyze the justifications of these assessments. This work provides insights into the characteristics that people consider human-like in the context of goal-directed video game navigation, which is a key step for further improving human interactions with AI agents.

Collaborative 3D object detection exploits information exchange among multiple agents to enhance accuracy of object detection in presence of sensor impairments such as occlusion. However, in practice, pose estimation errors due to imperfect localization would cause spatial message misalignment and significantly reduce the performance of collaboration. To alleviate adverse impacts of pose errors, we propose CoAlign, a novel hybrid collaboration framework that is robust to unknown pose errors. The proposed solution relies on a novel agent-object pose graph modeling to enhance pose consistency among collaborating agents. Furthermore, we adopt a multi-scale data fusion strategy to aggregate intermediate features at multiple spatial resolutions. Comparing with previous works, which require ground-truth pose for training supervision, our proposed CoAlign is more practical since it doesn't require any ground-truth pose supervision in the training and makes no specific assumptions on pose errors. Extensive evaluation of the proposed method is carried out on multiple datasets, certifying that CoAlign significantly reduce relative localization error and achieving the state of art detection performance when pose errors exist. Code are made available for the use of the research community at https://github.com/yifanlu0227/CoAlign.

Vision-and-Language Navigation (VLN) aims to develop intelligent agents to navigate in unseen environments only through language and vision supervision. In the recently proposed continuous settings (continuous VLN), the agent must act in a free 3D space and faces tougher challenges like real-time execution, complex instruction understanding, and long action sequence prediction. For a better performance in continuous VLN, we design a multi-level instruction understanding procedure and propose a novel model, Multi-Level Attention Network (MLANet). The first step of MLANet is to generate sub-instructions efficiently. We design a Fast Sub-instruction Algorithm (FSA) to segment the raw instruction into sub-instructions and generate a new sub-instruction dataset named ``FSASub". FSA is annotation-free and faster than the current method by 70 times, thus fitting the real-time requirement in continuous VLN. To solve the complex instruction understanding problem, MLANet needs a global perception of the instruction and observations. We propose a Multi-Level Attention (MLA) module to fuse vision, low-level semantics, and high-level semantics, which produce features containing a dynamic and global comprehension of the task. MLA also mitigates the adverse effects of noise words, thus ensuring a robust understanding of the instruction. To correctly predict actions in long trajectories, MLANet needs to focus on what sub-instruction is being executed every step. We propose a Peak Attention Loss (PAL) to improve the flexible and adaptive selection of the current sub-instruction. PAL benefits the navigation agent by concentrating its attention on the local information, thus helping the agent predict the most appropriate actions. We train and test MLANet in the standard benchmark. Experiment results show MLANet outperforms baselines by a significant margin.

We focus on the problem of rearranging a set of objects with a team of car-like robot pushers built using off-the-shelf components. Maintaining control of pushed objects while avoiding collisions in a tight space demands highly coordinated motion that is challenging to execute on constrained hardware. Centralized replanning approaches become intractable even for small-sized problems whereas decentralized approaches often get stuck in deadlocks. Our key insight is that by carefully assigning pushing tasks to robots, we could reduce the complexity of the rearrangement task, enabling robust performance via scalable decentralized control. Based on this insight, we built PuSHR, a system that optimally assigns pushing tasks and trajectories to robots offline, and performs trajectory tracking via decentralized control online. Through an ablation study in simulation, we demonstrate that PuSHR dominates baselines ranging from purely decentralized to fully decentralized in terms of success rate and time efficiency across challenging tasks with up to 4 robots. Hardware experiments demonstrate the transfer of our system to the real world and highlight its robustness to model inaccuracies. Our code can be found at https://github.com/prl-mushr/pushr, and videos from our experiments at https://youtu.be/DIWmZerF_O8.

Human-AI shared control allows human to interact and collaborate with autonomous agents to accomplish control tasks in complex environments. Previous Reinforcement Learning (RL) methods attempted goal-conditioned designs to achieve human-controllable policies at the cost of redesigning the reward function and training paradigm. Inspired by the neuroscience approach to investigate the motor cortex in primates, we develop a simple yet effective frequency-based approach called Policy Dissection to align the intermediate representation of the learned neural controller with the kinematic attributes of the agent behavior. Without modifying the neural controller or retraining the model, the proposed approach can convert a given RL-trained policy into a human-controllable policy. We evaluate the proposed approach on many RL tasks such as autonomous driving and locomotion. The experiments show that human-AI shared control system achieved by Policy Dissection in driving task can substantially improve the performance and safety in unseen traffic scenes. With human in the inference loop, the locomotion robots also exhibit versatile controllable motion skills even though they are only trained to move forward. Our results suggest the promising direction of implementing human-AI shared autonomy through interpreting the learned representation of the autonomous agents. Code and demo videos are available at https://metadriverse.github.io/policydissect.

In order for agents in multi-agent systems (MAS) to be safe, they need to take into account the risks posed by the actions of other agents. However, the dominant paradigm in game theory (GT) assumes that agents are not affected by risk from other agents and only strive to maximise their expected utility. For example, in hybrid human-AI driving systems, it is necessary to limit large deviations in reward resulting from car crashes. Although there are equilibrium concepts in game theory that take into account risk aversion, they either assume that agents are risk-neutral with respect to the uncertainty caused by the actions of other agents, or they are not guaranteed to exist. We introduce a new GT-based Risk-Averse Equilibrium (RAE) that always produces a solution that minimises the potential variance in reward accounting for the strategy of other agents. Theoretically and empirically, we show RAE shares many properties with a Nash Equilibrium (NE), establishing convergence properties and generalising to risk-dominant NE in certain cases. To tackle large-scale problems, we extend RAE to the PSRO multi-agent reinforcement learning (MARL) framework. We empirically demonstrate the minimum reward variance benefits of RAE in matrix games with high-risk outcomes. Results on MARL experiments show RAE generalises to risk-dominant NE in a trust dilemma game and that it reduces instances of crashing by 7x in an autonomous driving setting versus the best performing baseline.

We present a methodology to automatically compute worst-case performance bounds for a large class of first-order decentralized optimization algorithms. These algorithms aim at minimizing the average of local functions that are distributed across a network of agents. They typically combine local computations and consensus steps. Our methodology is based on the approach of Performance Estimation Problem (PEP), which allows computing the worst-case performance and a worst-case instance of first-order optimization algorithms by solving an SDP. We propose two ways of representing consensus steps in PEPs, which allow writing and solving PEPs for decentralized optimization. The first formulation is exact but specific to a given averaging matrix. The second formulation is a relaxation but provides guarantees valid over an entire class of averaging matrices, characterized by their spectral range. This formulation often allows recovering a posteriori the worst possible averaging matrix for the given algorithm. We apply our methodology to three different decentralized methods. For each of them, we obtain numerically tight worst-case performance bounds that significantly improve on the existing ones, as well as insights about the parameters tuning and the worst communication networks.