Nicholas R. Waytowich

Stephanie Milani, Anssi Kanervisto, Karolis Ramanauskas et al. · berkeley

To facilitate research in the direction of fine-tuning foundation models from human feedback, we held the MineRL BASALT Competition on Fine-Tuning from Human Feedback at NeurIPS 2022. The BASALT challenge asks teams to compete to develop algorithms to solve tasks with hard-to-specify reward functions in Minecraft. Through this competition, we aimed to promote the development of algorithms that use human feedback as channels to learn the desired behavior. We describe the competition and provide an overview of the top solutions. We conclude by discussing the impact of the competition and future directions for improvement.

Rohin Shah, Steven H. Wang, Cody Wild et al. · berkeley

We held the first-ever MineRL Benchmark for Agents that Solve Almost-Lifelike Tasks (MineRL BASALT) Competition at the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS 2021). The goal of the competition was to promote research towards agents that use learning from human feedback (LfHF) techniques to solve open-world tasks. Rather than mandating the use of LfHF techniques, we described four tasks in natural language to be accomplished in the video game Minecraft, and allowed participants to use any approach they wanted to build agents that could accomplish the tasks. Teams developed a diverse range of LfHF algorithms across a variety of possible human feedback types. The three winning teams implemented significantly different approaches while achieving similar performance. Interestingly, their approaches performed well on different tasks, validating our choice of tasks to include in the competition. While the outcomes validated the design of our competition, we did not get as many participants and submissions as our sister competition, MineRL Diamond. We speculate about the causes of this problem and suggest improvements for future iterations of the competition.

Vinicius G. Goecks, Nicholas R. Waytowich

The development of plans of action in disaster response scenarios is a time-consuming process. Large Language Models (LLMs) offer a powerful solution to expedite this process through in-context learning. This study presents DisasterResponseGPT, an algorithm that leverages LLMs to generate valid plans of action quickly by incorporating disaster response and planning guidelines in the initial prompt. In DisasterResponseGPT, users input the scenario description and receive a plan of action as output. The proposed method generates multiple plans within seconds, which can be further refined following the user's feedback. Preliminary results indicate that the plans of action developed by DisasterResponseGPT are comparable to human-generated ones while offering greater ease of modification in real-time. This approach has the potential to revolutionize disaster response operations by enabling rapid updates and adjustments during the plan's execution.

Bharat Prakash, Nicholas Waytowich, Tim Oates et al.

Learning to solve long horizon temporally extended tasks with reinforcement learning has been a challenge for several years now. We believe that it is important to leverage both the hierarchical structure of complex tasks and to use expert supervision whenever possible to solve such tasks. This work introduces an interpretable hierarchical agent framework by combining planning and semantic goal directed reinforcement learning. We assume access to certain spatial and haptic predicates and construct a simple and powerful semantic goal space. These semantic goal representations are more interpretable, making expert supervision and intervention easier. They also eliminate the need to write complex, dense reward functions thereby reducing human engineering effort. We evaluate our framework on a robotic block manipulation task and show that it performs better than other methods, including both sparse and dense reward functions. We also suggest some next steps and discuss how this framework makes interaction and collaboration with humans easier.

David Watkins, Peter Allen, Krzysztof Choromanski et al.

We propose the Multiple View Performer (MVP) - a new architecture for 3D shape completion from a series of temporally sequential views. MVP accomplishes this task by using linear-attention Transformers called Performers. Our model allows the current observation of the scene to attend to the previous ones for more accurate infilling. The history of past observations is compressed via the compact associative memory approximating modern continuous Hopfield memory, but crucially of size independent from the history length. We compare our model with several baselines for shape completion over time, demonstrating the generalization gains that MVP provides. To the best of our knowledge, MVP is the first multiple view voxel reconstruction method that does not require registration of multiple depth views and the first causal Transformer based model for 3D shape completion.

Nicholas R. Waytowich, Devin White, MD Sunbeam et al.

Recent advancements in large language models (LLMs) have expanded their capabilities beyond traditional text-based tasks to multimodal domains, integrating visual, auditory, and textual data. While multimodal LLMs have been extensively explored for high-level planning in domains like robotics and games, their potential as low-level controllers remains largely untapped. In this paper, we introduce a novel benchmark aimed at testing the emergent capabilities of multimodal LLMs as low-level policies in Atari games. Unlike traditional reinforcement learning (RL) methods that require training for each new environment and reward function specification, these LLMs utilize pre-existing multimodal knowledge to directly engage with game environments. Our study assesses the performances of multiple multimodal LLMs against traditional RL agents, human players, and random agents, focusing on their ability to understand and interact with complex visual scenes and formulate strategic responses. Our results show that these multimodal LLMs are not yet capable of being zero-shot low-level policies. Furthermore, we see that this is, in part, due to their visual and spatial reasoning. Additional results and videos are available on our project webpage: https://dev1nw.github.io/atari-gpt/.

Ellen Novoseller, Vinicius G. Goecks, David Watkins et al.

In machine learning for sequential decision-making, an algorithmic agent learns to interact with an environment while receiving feedback in the form of a reward signal. However, in many unstructured real-world settings, such a reward signal is unknown and humans cannot reliably craft a reward signal that correctly captures desired behavior. To solve tasks in such unstructured and open-ended environments, we present Demonstration-Inferred Preference Reinforcement Learning (DIP-RL), an algorithm that leverages human demonstrations in three distinct ways, including training an autoencoder, seeding reinforcement learning (RL) training batches with demonstration data, and inferring preferences over behaviors to learn a reward function to guide RL. We evaluate DIP-RL in a tree-chopping task in Minecraft. Results suggest that the method can guide an RL agent to learn a reward function that reflects human preferences and that DIP-RL performs competitively relative to baselines. DIP-RL is inspired by our previous work on combining demonstrations and pairwise preferences in Minecraft, which was awarded a research prize at the 2022 NeurIPS MineRL BASALT competition, Learning from Human Feedback in Minecraft. Example trajectory rollouts of DIP-RL and baselines are located at https://sites.google.com/view/dip-rl.

Aadi Palnitkar, Mingyang Mao, Nicholas Waytowich et al.

As large language models (LLMs) are applied to increasingly longer and more complex tasks, there is a growing need for realistic long-context benchmarks that require selective reading and integration of heterogeneous, multi-modal information sources. This need is especially acute for geospatial planning problems, such as those found in planning for large-scale military operations, which demand fast and accurate reasoning over maps, orders, intelligence reports, and other distributed data. To address this gap, we present MilSCORE (Military Scenario Contextual Reasoning), to our knowledge the first scenario-level dataset of expert-authored, multi-hop questions grounded in a complex, simulated military planning scenario used for training. MilSCORE is designed to evaluate high-stakes decision-making and planning, probing LLMs' ability to combine tactical and spatial reasoning across multiple sources and to reason over long-horizon, geospatially rich context. The benchmark includes a diverse set of question types across seven categories targeting both factual recall and multi-step reasoning about constraints, strategy, and spatial analysis. We provide an evaluation protocol and report baseline results for a range of contemporary vision-language models. Our findings highlight substantial headroom on MilSCORE, indicating that current systems struggle with realistic, scenario-level long-context planning, and positioning MilSCORE as a challenging testbed for future work.

Nicholas Waytowich, James Hare, Vinicius G. Goecks et al.

Traditionally, learning from human demonstrations via direct behavior cloning can lead to high-performance policies given that the algorithm has access to large amounts of high-quality data covering the most likely scenarios to be encountered when the agent is operating. However, in real-world scenarios, expert data is limited and it is desired to train an agent that learns a behavior policy general enough to handle situations that were not demonstrated by the human expert. Another alternative is to learn these policies with no supervision via deep reinforcement learning, however, these algorithms require a large amount of computing time to perform well on complex tasks with high-dimensional state and action spaces, such as those found in StarCraft II. Automatic curriculum learning is a recent mechanism comprised of techniques designed to speed up deep reinforcement learning by adjusting the difficulty of the current task to be solved according to the agent's current capabilities. Designing a proper curriculum, however, can be challenging for sufficiently complex tasks, and thus we leverage human demonstrations as a way to guide agent exploration during training. In this work, we aim to train deep reinforcement learning agents that can command multiple heterogeneous actors where starting positions and overall difficulty of the task are controlled by an automatically-generated curriculum from a single human demonstration. Our results show that an agent trained via automated curriculum learning can outperform state-of-the-art deep reinforcement learning baselines and match the performance of the human expert in a simulated command and control task in StarCraft II modeled over a real military scenario.

Mingyang Mao, Mariela M. Perez-Cabarcas, Utteja Kallakuri et al.

To effectively engage in human society, the ability to adapt, filter information, and make informed decisions in ever-changing situations is critical. As robots and intelligent agents become more integrated into human life, there is a growing opportunity-and need-to offload the cognitive burden on humans to these systems, particularly in dynamic, information-rich scenarios. To fill this critical need, we present Multi-RAG, a multimodal retrieval-augmented generation system designed to provide adaptive assistance to humans in information-intensive circumstances. Our system aims to improve situational understanding and reduce cognitive load by integrating and reasoning over multi-source information streams, including video, audio, and text. As an enabling step toward long-term human-robot partnerships, Multi-RAG explores how multimodal information understanding can serve as a foundation for adaptive robotic assistance in dynamic, human-centered situations. To evaluate its capability in a realistic human-assistance proxy task, we benchmarked Multi-RAG on the MMBench-Video dataset, a challenging multimodal video understanding benchmark. Our system achieves superior performance compared to existing open-source video large language models (Video-LLMs) and large vision-language models (LVLMs), while utilizing fewer resources and less input data. The results demonstrate Multi- RAG's potential as a practical and efficient foundation for future human-robot adaptive assistance systems in dynamic, real-world contexts.

Jonathan Hyun, Nicholas R Waytowich, Boyuan Chen

Despite rapid progress in large language model (LLM)-based multi-agent systems, current benchmarks fall short in evaluating their scalability, robustness, and coordination capabilities in complex, dynamic, real-world tasks. Existing environments typically focus on small-scale, fully observable, or low-complexity domains, limiting their utility for developing and assessing next-generation multi-agent Agentic AI frameworks. We introduce CREW-Wildfire, an open-source benchmark designed to close this gap. Built atop the human-AI teaming CREW simulation platform, CREW-Wildfire offers procedurally generated wildfire response scenarios featuring large maps, heterogeneous agents, partial observability, stochastic dynamics, and long-horizon planning objectives. The environment supports both low-level control and high-level natural language interactions through modular Perception and Execution modules. We implement and evaluate several state-of-the-art LLM-based multi-agent Agentic AI frameworks, uncovering significant performance gaps that highlight the unsolved challenges in large-scale coordination, communication, spatial reasoning, and long-horizon planning under uncertainty. By providing more realistic complexity, scalable architecture, and behavioral evaluation metrics, CREW-Wildfire establishes a critical foundation for advancing research in scalable multi-agent Agentic intelligence. All code, environments, data, and baselines will be released to support future research in this emerging domain.

Vinicius G. Goecks, Nicholas Waytowich, David Watkins-Valls et al.

Real-world tasks of interest are generally poorly defined by human-readable descriptions and have no pre-defined reward signals unless it is defined by a human designer. Conversely, data-driven algorithms are often designed to solve a specific, narrowly defined, task with performance metrics that drives the agent's learning. In this work, we present the solution that won first place and was awarded the most human-like agent in the 2021 NeurIPS Competition MineRL BASALT Challenge: Learning from Human Feedback in Minecraft, which challenged participants to use human data to solve four tasks defined only by a natural language description and no reward function. Our approach uses the available human demonstration data to train an imitation learning policy for navigation and additional human feedback to train an image classifier. These modules, combined with an estimated odometry map, become a powerful state-machine designed to utilize human knowledge in a natural hierarchical paradigm. We compare this hybrid intelligence approach to both end-to-end machine learning and pure engineered solutions, which are then judged by human evaluators. Codebase is available at https://github.com/viniciusguigo/kairos_minerl_basalt.

Vinicius G. Goecks, Nicholas Waytowich

The development of Courses of Action (COAs) in military operations is traditionally a time-consuming and intricate process. Addressing this challenge, this study introduces COA-GPT, a novel algorithm employing Large Language Models (LLMs) for rapid and efficient generation of valid COAs. COA-GPT incorporates military doctrine and domain expertise to LLMs through in-context learning, allowing commanders to input mission information - in both text and image formats - and receive strategically aligned COAs for review and approval. Uniquely, COA-GPT not only accelerates COA development, producing initial COAs within seconds, but also facilitates real-time refinement based on commander feedback. This work evaluates COA-GPT in a military-relevant scenario within a militarized version of the StarCraft II game, comparing its performance against state-of-the-art reinforcement learning algorithms. Our results demonstrate COA-GPT's superiority in generating strategically sound COAs more swiftly, with added benefits of enhanced adaptability and alignment with commander intentions. COA-GPT's capability to rapidly adapt and update COAs during missions presents a transformative potential for military planning, particularly in addressing planning discrepancies and capitalizing on emergent windows of opportunities.

Anna Madison, Ellen Novoseller, Vinicius G. Goecks et al.

Future warfare will require Command and Control (C2) personnel to make decisions at shrinking timescales in complex and potentially ill-defined situations. Given the need for robust decision-making processes and decision-support tools, integration of artificial and human intelligence holds the potential to revolutionize the C2 operations process to ensure adaptability and efficiency in rapidly changing operational environments. We propose to leverage recent promising breakthroughs in interactive machine learning, in which humans can cooperate with machine learning algorithms to guide machine learning algorithm behavior. This paper identifies several gaps in state-of-the-art science and technology that future work should address to extend these approaches to function in complex C2 contexts. In particular, we describe three research focus areas that together, aim to enable scalable interactive machine learning (SIML): 1) developing human-AI interaction algorithms to enable planning in complex, dynamic situations; 2) fostering resilient human-AI teams through optimizing roles, configurations, and trust; and 3) scaling algorithms and human-AI teams for flexibility across a range of potential contexts and situations.

Sean Kulinski, Nicholas R. Waytowich, James Z. Hare et al.

Spatial reasoning tasks in multi-agent environments such as event prediction, agent type identification, or missing data imputation are important for multiple applications (e.g., autonomous surveillance over sensor networks and subtasks for reinforcement learning (RL)). StarCraft II game replays encode intelligent (and adversarial) multi-agent behavior and could provide a testbed for these tasks; however, extracting simple and standardized representations for prototyping these tasks is laborious and hinders reproducibility. In contrast, MNIST and CIFAR10, despite their extreme simplicity, have enabled rapid prototyping and reproducibility of ML methods. Following the simplicity of these datasets, we construct a benchmark spatial reasoning dataset based on StarCraft II replays that exhibit complex multi-agent behaviors, while still being as easy to use as MNIST and CIFAR10. Specifically, we carefully summarize a window of 255 consecutive game states to create 3.6 million summary images from 60,000 replays, including all relevant metadata such as game outcome and player races. We develop three formats of decreasing complexity: Hyperspectral images that include one channel for every unit type (similar to multispectral geospatial images), RGB images that mimic CIFAR10, and grayscale images that mimic MNIST. We show how this dataset can be used for prototyping spatial reasoning methods. All datasets, code for extraction, and code for dataset loading can be found at https://starcraftdata.davidinouye.com

Mikolaj Walczak, Romina Aalishah, Wyatt Mackey et al.

Deep reinforcement learning agents are often fragile while humans remain adaptive and flexible to varying scenarios. To bridge this gap, we present EDEN, a biologically inspired navigation framework that integrates learned entorhinal-like grid cell representations and reinforcement learning to enable autonomous navigation. Inspired by the mammalian entorhinal-hippocampal system, EDEN allows agents to perform path integration and vector-based navigation using visual and motion sensor data. At the core of EDEN is a grid cell encoder that transforms egocentric motion into periodic spatial codes, producing low-dimensional, interpretable embeddings of position. To generate these activations from raw sensory input, we combine fiducial marker detections in the lightweight MiniWorld simulator and DINO-based visual features in the high-fidelity Gazebo simulator. These spatial representations serve as input to a policy trained with Proximal Policy Optimization (PPO), enabling dynamic, goal-directed navigation. We evaluate EDEN in both MiniWorld, for rapid prototyping, and Gazebo, which offers realistic physics and perception noise. Compared to baseline agents using raw state inputs (e.g., position, velocity) or standard convolutional image encoders, EDEN achieves a 99% success rate, within the simple scenarios, and >94% within complex floorplans with occluded paths with more efficient and reliable step-wise navigation. In addition, as a replacement of ground truth activations, we present a trainable Grid Cell encoder enabling the development of periodic grid-like patterns from vision and motion sensor data, emulating the development of such patterns within biological mammals. This work represents a step toward biologically grounded spatial intelligence in robotics, bridging neural navigation principles with reinforcement learning for scalable deployment.

Evelyn Rose, Devin White, Mingkang Wu et al.

This paper explores multiple optimization methods to improve the performance of rating-based reinforcement learning (RbRL). RbRL, a method based on the idea of human ratings, has been developed to infer reward functions in reward-free environments for the subsequent policy learning via standard reinforcement learning, which requires the availability of reward functions. Specifically, RbRL minimizes the cross entropy loss that quantifies the differences between human ratings and estimated ratings derived from the inferred reward. Hence, a low loss means a high degree of consistency between human ratings and estimated ratings. Despite its simple form, RbRL has various hyperparameters and can be sensitive to various factors. Therefore, it is critical to provide comprehensive experiments to understand the impact of various hyperparameters on the performance of RbRL. This paper is a work in progress, providing users some general guidelines on how to select hyperparameters in RbRL.

Connor Mattson, Varun Raveendra, Ellen Novoseller et al.

Imitation Learning (IL) is a natural way for humans to teach robots, particularly when high-quality demonstrations are easy to obtain. While IL has been widely applied to single-robot settings, relatively few studies have addressed the extension of these methods to multi-agent systems, especially in settings where a single human must provide demonstrations to a team of collaborating robots. In this paper, we introduce and study Round-Robin Behavior Cloning (R2BC), a method that enables a single human operator to effectively train multi-robot systems through sequential, single-agent demonstrations. Our approach allows the human to teleoperate one agent at a time and incrementally teach multi-agent behavior to the entire system, without requiring demonstrations in the joint multi-agent action space. We show that R2BC methods match, and in some cases surpass, the performance of an oracle behavior cloning approach trained on privileged synchronized demonstrations across four multi-agent simulated tasks. Finally, we deploy R2BC on two physical robot tasks trained using real human demonstrations.

Mingkang Wu, Devin White, Evelyn Rose et al.

Reinforcement learning from human feedback (RLHF) has become a key factor in aligning model behavior with users' goals. However, while humans integrate multiple strategies when making decisions, current RLHF approaches often simplify this process by modeling human reasoning through isolated tasks such as classification or regression. In this paper, we propose a novel reinforcement learning (RL) method that mimics human decision-making by jointly considering multiple tasks. Specifically, we leverage human ratings in reward-free environments to infer a reward function, introducing learnable weights that balance the contributions of both classification and regression models. This design captures the inherent uncertainty in human decision-making and allows the model to adaptively emphasize different strategies. We conduct several experiments using synthetic human ratings to validate the effectiveness of the proposed approach. Results show that our method consistently outperforms existing rating-based RL methods, and in some cases, even surpasses traditional RL approaches.

Bharat Prakash, Nicholas Waytowich, Tinoosh Mohsenin et al.

Reinforcement Learning (RL) agents can learn to solve complex sequential decision making tasks by interacting with the environment. However, sample efficiency remains a major challenge. In the field of multi-goal RL, where agents are required to reach multiple goals to solve complex tasks, improving sample efficiency can be especially challenging. On the other hand, humans or other biological agents learn such tasks in a much more strategic way, following a curriculum where tasks are sampled with increasing difficulty level in order to make gradual and efficient learning progress. In this work, we propose a method for automatic goal generation using a dynamical distance function (DDF) in a self-supervised fashion. DDF is a function which predicts the dynamical distance between any two states within a markov decision process (MDP). With this, we generate a curriculum of goals at the appropriate difficulty level to facilitate efficient learning throughout the training process. We evaluate this approach on several goal-conditioned robotic manipulation and navigation tasks, and show improvements in sample efficiency over a baseline method which only uses random goal sampling.

Vinicius G. Goecks, Nicholas Waytowich, Derrik E. Asher et al.

Games and simulators can be a valuable platform to execute complex multi-agent, multiplayer, imperfect information scenarios with significant parallels to military applications: multiple participants manage resources and make decisions that command assets to secure specific areas of a map or neutralize opposing forces. These characteristics have attracted the artificial intelligence (AI) community by supporting development of algorithms with complex benchmarks and the capability to rapidly iterate over new ideas. The success of artificial intelligence algorithms in real-time strategy games such as StarCraft II have also attracted the attention of the military research community aiming to explore similar techniques in military counterpart scenarios. Aiming to bridge the connection between games and military applications, this work discusses past and current efforts on how games and simulators, together with the artificial intelligence algorithms, have been adapted to simulate certain aspects of military missions and how they might impact the future battlefield. This paper also investigates how advances in virtual reality and visual augmentation systems open new possibilities in human interfaces with gaming platforms and their military parallels.

Bharat Prakash, Nicholas Waytowich, Tim Oates et al.

Reinforcement learning has been successful in many tasks ranging from robotic control, games, energy management etc. In complex real world environments with sparse rewards and long task horizons, sample efficiency is still a major challenge. Most complex tasks can be easily decomposed into high-level planning and low level control. Therefore, it is important to enable agents to leverage the hierarchical structure and decompose bigger tasks into multiple smaller sub-tasks. We introduce an approach where we use language to specify sub-tasks and a high-level planner issues language commands to a low level controller. The low-level controller executes the sub-tasks based on the language commands. Our experiments show that this method is able to solve complex long horizon planning tasks with limited human supervision. Using language has added benefit of interpretability and ability for expert humans to take over the high-level planning task and provide language commands if necessary.

David Watkins, Peter K Allen, Henrique Maia et al.

While both navigation and manipulation are challenging topics in isolation, many tasks require the ability to both navigate and manipulate in concert. To this end, we propose a mobile manipulation system that leverages novel navigation and shape completion methods to manipulate an object with a mobile robot. Our system utilizes uncertainty in the initial estimation of a manipulation target to calculate a predicted next-best-view. Without the need of localization, the robot then uses the predicted panoramic view at the next-best-view location to navigate to the desired location, capture a second view of the object, create a new model that predicts the shape of object more accurately than a single image alone, and uses this model for grasp planning. We show that the system is highly effective for mobile manipulation tasks through simulation experiments using real world data, as well as ablations on each component of our system.

Ravi Kumar Thakur, MD-Nazmus Samin Sunbeam, Vinicius G. Goecks et al.

Approaches for teaching learning agents via human demonstrations have been widely studied and successfully applied to multiple domains. However, the majority of imitation learning work utilizes only behavioral information from the demonstrator, i.e. which actions were taken, and ignores other useful information. In particular, eye gaze information can give valuable insight towards where the demonstrator is allocating visual attention, and holds the potential to improve agent performance and generalization. In this work, we propose Gaze Regularized Imitation Learning (GRIL), a novel context-aware, imitation learning architecture that learns concurrently from both human demonstrations and eye gaze to solve tasks where visual attention provides important context. We apply GRIL to a visual navigation task, in which an unmanned quadrotor is trained to search for and navigate to a target vehicle in a photorealistic simulated environment. We show that GRIL outperforms several state-of-the-art gaze-based imitation learning algorithms, simultaneously learns to predict human visual attention, and generalizes to scenarios not present in the training data. Supplemental videos and code can be found at https://sites.google.com/view/gaze-regularized-il/.

Ritwik Bera, Vinicius G. Goecks, Gregory M. Gremillion et al.

Learning from demonstration has been widely studied in machine learning but becomes challenging when the demonstrated trajectories are unstructured and follow different objectives. This short-paper proposes PODNet, Plannable Option Discovery Network, addressing how to segment an unstructured set of demonstrated trajectories for option discovery. This enables learning from demonstration to perform multiple tasks and plan high-level trajectories based on the discovered option labels. PODNet combines a custom categorical variational autoencoder, a recurrent option inference network, option-conditioned policy network, and option dynamics model in an end-to-end learning architecture. Due to the concurrently trained option-conditioned policy network and option dynamics model, the proposed architecture has implications in multi-task and hierarchical learning, explainable and interpretable artificial intelligence, and applications where the agent is required to learn only from observations.

Nicholas Waytowich, Sean L. Barton, Vernon Lawhern et al.

While deep reinforcement learning techniques have led to agents that are successfully able to learn to perform a number of tasks that had been previously unlearnable, these techniques are still susceptible to the longstanding problem of reward sparsity. This is especially true for tasks such as training an agent to play StarCraft II, a real-time strategy game where reward is only given at the end of a game which is usually very long. While this problem can be addressed through reward shaping, such approaches typically require a human expert with specialized knowledge. Inspired by the vision of enabling reward shaping through the more-accessible paradigm of natural-language narration, we develop a technique that can provide the benefits of reward shaping using natural language commands. Our narration-guided RL agent projects sequences of natural-language commands into the same high-dimensional representation space as corresponding goal states. We show that we can get improved performance with our method compared to traditional reward-shaping approaches. Additionally, we demonstrate the ability of our method to generalize to unseen natural-language commands.

Vinicius G. Goecks, Gregory M. Gremillion, Vernon J. Lawhern et al.

This paper investigates how to efficiently transition and update policies, trained initially with demonstrations, using off-policy actor-critic reinforcement learning. It is well-known that techniques based on Learning from Demonstrations, for example behavior cloning, can lead to proficient policies given limited data. However, it is currently unclear how to efficiently update that policy using reinforcement learning as these approaches are inherently optimizing different objective functions. Previous works have used loss functions, which combine behavior cloning losses with reinforcement learning losses to enable this update. However, the components of these loss functions are often set anecdotally, and their individual contributions are not well understood. In this work, we propose the Cycle-of-Learning (CoL) framework that uses an actor-critic architecture with a loss function that combines behavior cloning and 1-step Q-learning losses with an off-policy pre-training step from human demonstrations. This enables transition from behavior cloning to reinforcement learning without performance degradation and improves reinforcement learning in terms of overall performance and training time. Additionally, we carefully study the composition of these combined losses and their impact on overall policy learning. We show that our approach outperforms state-of-the-art techniques for combining behavior cloning and reinforcement learning for both dense and sparse reward scenarios. Our results also suggest that directly including the behavior cloning loss on demonstration data helps to ensure stable learning and ground future policy updates.

Sunil Gandhi, Tim Oates, Tinoosh Mohsenin et al.

In this paper, we present a method for learning from video demonstrations by using human feedback to construct a mapping between the standard representation of the agent and the visual representation of the demonstration. In this way, we leverage the advantages of both these representations, i.e., we learn the policy using standard state representations, but are able to specify the expected behavior using video demonstration. We train an autonomous agent using a single video demonstration and use human feedback (using numerical similarity rating) to map the standard representation to the visual representation with a neural network. We show the effectiveness of our method by teaching a hopper agent in the MuJoCo to perform a backflip using a single video demonstration generated in MuJoCo as well as from a real-world YouTube video of a person performing a backflip. Additionally, we show that our method can transfer to new tasks, such as hopping, with very little human feedback.

David Watkins-Valls, Jingxi Xu, Nicholas Waytowich et al.

We present a robot navigation system that uses an imitation learning framework to successfully navigate in complex environments. Our framework takes a pre-built 3D scan of a real environment and trains an agent from pre-generated expert trajectories to navigate to any position given a panoramic view of the goal and the current visual input without relying on map, compass, odometry, or relative position of the target at runtime. Our end-to-end trained agent uses RGB and depth (RGBD) information and can handle large environments (up to $1031m^2$) across multiple rooms (up to $40$) and generalizes to unseen targets. We show that when compared to several baselines our method (1) requires fewer training examples and less training time, (2) reaches the goal location with higher accuracy, and (3) produces better solutions with shorter paths for long-range navigation tasks.

Pietro Pierpaoli, Harish Ravichandar, Nicholas Waytowich et al.

We present a technique for learning how to solve a multi-robot mission that requires interaction with an external environment by observing an expert system executing the same mission. We define the expert system as a team of robots equipped with a library of controllers, each designed to solve a specific task, supervised by an expert policy that appropriately selects controllers based on the states of robots and environment. The objective is for an un-trained team of robots (i.e., imitator system) equipped with the same library of controllers, but agnostic to the expert policy, to execute the mission, with performances comparable to those of the expert system. From un-annotated observations of the expert system, a multi-hypothesis filtering technique is used to estimate individual controllers executed by the expert policy. Then, the history of estimated controllers and environmental states is used to train a neural network policy for the imitator system. Considering a perimeter protection scenario on a team of differential-drive robots, we show that the learned policy endows the imitator system with performances comparable to those of the expert system.

Yilun Zhou, Derrik E. Asher, Nicholas R. Waytowich et al.

Multi-agent reinforcement learning (MARL) extends (single-agent) reinforcement learning (RL) by introducing additional agents and (potentially) partial observability of the environment. Consequently, algorithms for solving MARL problems incorporate various extensions beyond traditional RL methods, such as a learned communication protocol between cooperative agents that enables exchange of private information or adaptive modeling of opponents in competitive settings. One popular algorithmic construct is a memory mechanism such that an agent's decisions can depend not only upon the current state but also upon the history of observed states and actions. In this paper, we study how a memory mechanism can be useful in environments with different properties, such as observability, internality and presence of a communication channel. Using both prior work and new experiments, we show that a memory mechanism is helpful when learning agents need to model other agents and/or when communication is constrained in some way; however we must to be cautious of agents achieving effective memoryfulness through other means.

Nicholas Waytowich, Sean L. Barton, Vernon Lawhern et al.

While deep reinforcement learning techniques have led to agents that are successfully able to learn to perform a number of tasks that had been previously unlearnable, these techniques are still susceptible to the longstanding problem of {\em reward sparsity}. This is especially true for tasks such as training an agent to play StarCraft II, a real-time strategy game where reward is only given at the end of a game which is usually very long. While this problem can be addressed through reward shaping, such approaches typically require a human expert with specialized knowledge. Inspired by the vision of enabling reward shaping through the more-accessible paradigm of natural-language narration, we investigate to what extent we can contextualize these narrations by grounding them to the goal-specific states. We present a mutual-embedding model using a multi-input deep-neural network that projects a sequence of natural language commands into the same high-dimensional representation space as corresponding goal states. We show that using this model we can learn an embedding space with separable and distinct clusters that accurately maps natural-language commands to corresponding game states . We also discuss how this model can allow for the use of narrations as a robust form of reward shaping to improve RL performance and efficiency.

Bharat Prakash, Mark Horton, Nicholas R. Waytowich et al.

In autonomous embedded systems, it is often vital to reduce the amount of actions taken in the real world and energy required to learn a policy. Training reinforcement learning agents from high dimensional image representations can be very expensive and time consuming. Autoencoders are deep neural network used to compress high dimensional data such as pixelated images into small latent representations. This compression model is vital to efficiently learn policies, especially when learning on embedded systems. We have implemented this model on the NVIDIA Jetson TX2 embedded GPU, and evaluated the power consumption, throughput, and energy consumption of the autoencoders for various CPU/GPU core combinations, frequencies, and model parameters. Additionally, we have shown the reconstructions generated by the autoencoder to analyze the quality of the generated compressed representation and also the performance of the reinforcement learning agent. Finally, we have presented an assessment of the viability of training these models on embedded systems and their usefulness in developing autonomous policies. Using autoencoders, we were able to achieve 4-5 $\times$ improved performance compared to a baseline RL agent with a convolutional feature extractor, while using less than 2W of power.

Bharat Prakash, Mohit Khatwani, Nicholas Waytowich et al.

Recent progress in AI and Reinforcement learning has shown great success in solving complex problems with high dimensional state spaces. However, most of these successes have been primarily in simulated environments where failure is of little or no consequence. Most real-world applications, however, require training solutions that are safe to operate as catastrophic failures are inadmissible especially when there is human interaction involved. Currently, Safe RL systems use human oversight during training and exploration in order to make sure the RL agent does not go into a catastrophic state. These methods require a large amount of human labor and it is very difficult to scale up. We present a hybrid method for reducing the human intervention time by combining model-based approaches and training a supervised learner to improve sample efficiency while also ensuring safety. We evaluate these methods on various grid-world environments using both standard and visual representations and show that our approach achieves better performance in terms of sample efficiency, number of catastrophic states reached as well as overall task performance compared to traditional model-free approaches

Vinicius G. Goecks, Gregory M. Gremillion, Vernon J. Lawhern et al.

This paper investigates how to utilize different forms of human interaction to safely train autonomous systems in real-time by learning from both human demonstrations and interventions. We implement two components of the Cycle-of-Learning for Autonomous Systems, which is our framework for combining multiple modalities of human interaction. The current effort employs human demonstrations to teach a desired behavior via imitation learning, then leverages intervention data to correct for undesired behaviors produced by the imitation learner to teach novel tasks to an autonomous agent safely, after only minutes of training. We demonstrate this method in an autonomous perching task using a quadrotor with continuous roll, pitch, yaw, and throttle commands and imagery captured from a downward-facing camera in a high-fidelity simulated environment. Our method improves task completion performance for the same amount of human interaction when compared to learning from demonstrations alone, while also requiring on average 32% less data to achieve that performance. This provides evidence that combining multiple modes of human interaction can increase both the training speed and overall performance of policies for autonomous systems.

Sean L. Barton, Nicholas R. Waytowich, Derrik E. Asher

We discuss the role of coordination as a direct learning objective in multi-agent reinforcement learning (MARL) domains. To this end, we present a novel means of quantifying coordination in multi-agent systems, and discuss the implications of using such a measure to optimize coordinated agent policies. This concept has important implications for adversary-aware RL, which we take to be a sub-domain of multi-agent learning.

Nicholas R. Waytowich, Vinicius G. Goecks, Vernon J. Lawhern

We discuss different types of human-robot interaction paradigms in the context of training end-to-end reinforcement learning algorithms. We provide a taxonomy to categorize the types of human interaction and present our Cycle-of-Learning framework for autonomous systems that combines different human-interaction modalities with reinforcement learning. Two key concepts provided by our Cycle-of-Learning framework are how it handles the integration of the different human-interaction modalities (demonstration, intervention, and evaluation) and how to define the switching criteria between them.

Nicholas R. Waytowich, Vernon Lawhern, Javier O. Garcia et al.

Steady-State Visual Evoked Potentials (SSVEPs) are neural oscillations from the parietal and occipital regions of the brain that are evoked from flickering visual stimuli. SSVEPs are robust signals measurable in the electroencephalogram (EEG) and are commonly used in brain-computer interfaces (BCIs). However, methods for high-accuracy decoding of SSVEPs usually require hand-crafted approaches that leverage domain-specific knowledge of the stimulus signals, such as specific temporal frequencies in the visual stimuli and their relative spatial arrangement. When this knowledge is unavailable, such as when SSVEP signals are acquired asynchronously, such approaches tend to fail. In this paper, we show how a compact convolutional neural network (Compact-CNN), which only requires raw EEG signals for automatic feature extraction, can be used to decode signals from a 12-class SSVEP dataset without the need for any domain-specific knowledge or calibration data. We report across subject mean accuracy of approximately 80% (chance being 8.3%) and show this is substantially better than current state-of-the-art hand-crafted approaches using canonical correlation analysis (CCA) and Combined-CCA. Furthermore, we analyze our Compact-CNN to examine the underlying feature representation, discovering that the deep learner extracts additional phase and amplitude related features associated with the structure of the dataset. We discuss how our Compact-CNN shows promise for BCI applications that allow users to freely gaze/attend to any stimulus at any time (e.g., asynchronous BCI) as well as provides a method for analyzing SSVEP signals in a way that might augment our understanding about the basic processing in the visual cortex.

Garrett Warnell, Nicholas Waytowich, Vernon Lawhern et al.

While recent advances in deep reinforcement learning have allowed autonomous learning agents to succeed at a variety of complex tasks, existing algorithms generally require a lot of training data. One way to increase the speed at which agents are able to learn to perform tasks is by leveraging the input of human trainers. Although such input can take many forms, real-time, scalar-valued feedback is especially useful in situations where it proves difficult or impossible for humans to provide expert demonstrations. Previous approaches have shown the usefulness of human input provided in this fashion (e.g., the TAMER framework), but they have thus far not considered high-dimensional state spaces or employed the use of deep learning. In this paper, we do both: we propose Deep TAMER, an extension of the TAMER framework that leverages the representational power of deep neural networks in order to learn complex tasks in just a short amount of time with a human trainer. We demonstrate Deep TAMER's success by using it and just 15 minutes of human-provided feedback to train an agent that performs better than humans on the Atari game of Bowling - a task that has proven difficult for even state-of-the-art reinforcement learning methods.