Alan Fern

Jeremy Dao, Kevin Green, Helei Duan et al.

Recent work on sim-to-real learning for bipedal locomotion has demonstrated new levels of robustness and agility over a variety of terrains. However, that work, and most prior bipedal locomotion work, have not considered locomotion under a variety of external loads that can significantly influence the overall system dynamics. In many applications, robots will need to maintain robust locomotion under a wide range of potential dynamic loads, such as pulling a cart or carrying a large container of sloshing liquid, ideally without requiring additional load-sensing capabilities. In this work, we explore the capabilities of reinforcement learning (RL) and sim-to-real transfer for bipedal locomotion under dynamic loads using only proprioceptive feedback. We show that prior RL policies trained for unloaded locomotion fail for some loads and that simply training in the context of loads is enough to result in successful and improved policies. We also compare training specialized policies for each load versus a single policy for all considered loads and analyze how the resulting gaits change to accommodate different loads. Finally, we demonstrate sim-to-real transfer, which is successful but shows a wider sim-to-real gap than prior unloaded work, which points to interesting future research.

Rich Meier, Eduardo Cotilla-Sanchez, Alan Fern

In recent years there have been many improvements in the reliability of critical infrastructure systems. Despite these improvements, the power systems industry has seen relatively small advances in this regard. For instance, power quality deficiencies, a high number of localized contingencies, and large cascading outages are still too widespread. Though progress has been made in improving generation, transmission, and distribution infrastructure, remedial action schemes (RAS) remain non-standardized and are often not uniformly implemented across different utilities, ISOs, and RTOs. Traditionally, load shedding and islanding have been successful protection measures in restraining propagation of contingencies and large cascading outages. This paper proposes a novel, algorithmic approach to selecting RAS policies to optimize the operation of the power network during and after a contingency. Specifically, we use policy-switching to consolidate traditional load shedding and islanding schemes. In order to model and simulate the functionality of the proposed power systems protection algorithm, we conduct Monte-Carlo, time-domain simulations using Siemens PSS/E. The algorithm is tested via experiments on the IEEE-39 topology to demonstrate that the proposed approach achieves optimal power system performance during emergency situations, given a specific set of RAS policies.

Aseem Saxena, Paola Pesantez-Cabrera, Rohan Ballapragada et al.

Cold temperatures during fall and spring have the potential to cause frost damage to grapevines and other fruit plants, which can significantly decrease harvest yields. To help prevent these losses, farmers deploy expensive frost mitigation measures such as sprinklers, heaters, and wind machines when they judge that damage may occur. This judgment, however, is challenging because the cold hardiness of plants changes throughout the dormancy period and it is difficult to directly measure. This has led scientists to develop cold hardiness prediction models that can be tuned to different grape cultivars based on laborious field measurement data. In this paper, we study whether deep learning models can improve cold hardiness prediction for grapes based on data that has been collected over a 30-year time period. A key challenge is that the amount of data per cultivar is highly variable, with some cultivars having only a small amount. For this purpose, we investigate the use of multi-task learning to leverage data across cultivars in order to improve prediction performance for individual cultivars. We evaluate a number of multi-task learning approaches and show that the highest performing approach is able to significantly improve over learning for single cultivars and outperforms the current state-of-the-art scientific model for most cultivars.

Aseem Saxena, Paola Pesantez-Cabrera, Rohan Ballapragada et al.

Grapevine budbreak is a key phenological stage of seasonal development, which serves as a signal for the onset of active growth. This is also when grape plants are most vulnerable to damage from freezing temperatures. Hence, it is important for winegrowers to anticipate the day of budbreak occurrence to protect their vineyards from late spring frost events. This work investigates deep learning for budbreak prediction using data collected for multiple grape cultivars. While some cultivars have over 30 seasons of data others have as little as 4 seasons, which can adversely impact prediction accuracy. To address this issue, we investigate multi-task learning, which combines data across all cultivars to make predictions for individual cultivars. Our main result shows that several variants of multi-task learning are all able to significantly improve prediction accuracy compared to learning for each cultivar independently.

Kin-Ho Lam, Delyar Tabatabai, Jed Irvine et al.

Reinforcement learning (RL) agents are commonly evaluated via their expected value over a distribution of test scenarios. Unfortunately, this evaluation approach provides limited evidence for post-deployment generalization beyond the test distribution. In this paper, we address this limitation by extending the recent CheckList testing methodology from natural language processing to planning-based RL. Specifically, we consider testing RL agents that make decisions via online tree search using a learned transition model and value function. The key idea is to improve the assessment of future performance via a CheckList approach for exploring and assessing the agent's inferences during tree search. The approach provides the user with an interface and general query-rule mechanism for identifying potential inference flaws and validating expected inference invariances. We present a user study involving knowledgeable AI researchers using the approach to evaluate an agent trained to play a complex real-time strategy game. The results show the approach is effective in allowing users to identify previously-unknown flaws in the agent's reasoning. In addition, our analysis provides insight into how AI experts use this type of testing approach, which may help improve future instantiations.

Anurag Koul, Mariano Phielipp, Alan Fern

Decision makers often wish to use offline historical data to compare sequential-action policies at various world states. Importantly, computational tools should produce confidence values for such offline policy comparison (OPC) to account for statistical variance and limited data coverage. Nevertheless, there is little work that directly evaluates the quality of confidence values for OPC. In this work, we address this issue by creating benchmarks for OPC with Confidence (OPCC), derived by adding sets of policy comparison queries to datasets from offline reinforcement learning. In addition, we present an empirical evaluation of the risk versus coverage trade-off for a class of model-based baselines. In particular, the baselines learn ensembles of dynamics models, which are used in various ways to produce simulations for answering queries with confidence values. While our results suggest advantages for certain baseline variations, there appears to be significant room for improvement in future work.

William Solow, Paola Pesantez-Cabrera, Markus Keller et al.

Accurate prediction of crop states (e.g., phenology stages and cold hardiness) is essential for timely farm management decisions such as irrigation, fertilization, and canopy management to optimize crop yield and quality. While traditional biophysical models can be used for season-long predictions, they lack the precision required for site-specific management. Deep learning methods are a compelling alternative, but can produce biologically unrealistic predictions and require large-scale data. We propose a \emph{hybrid modeling} approach that uses a neural network to parameterize a differentiable biophysical model and leverages multi-task learning for efficient data sharing across crop cultivars in data limited settings. By predicting the \emph{parameters} of the biophysical model, our approach improves the prediction accuracy while preserving biological realism. Empirical evaluation using real-world and synthetic datasets demonstrates that our method improves prediction accuracy by 60\% for phenology and 40\% for cold hardiness compared to deployed biophysical models.

Jeremy Dao, Alan Fern

Simulation trained legged locomotion policies often exhibit performance loss on hardware due to dynamics discrepancies between the simulator and the real world, highlighting the need for approaches that adapt the simulator itself to better match hardware behavior. Prior work typically quantify these discrepancies through precise, time-aligned matching of joint and base trajectories. This process requires motion capture, privileged sensing, and carefully controlled initial conditions. We introduce a practical alternative based on proprioceptive distribution matching, which compares hardware and simulation rollouts as distributions of joint observations and actions, eliminating the need for time alignment or external sensing. Using this metric as a black-box objective, we explore adapting simulator dynamics through parameter identification, action-delta models, and residual actuator models. Our approach matches the parameter recovery and policy-performance gains of privileged state-matching baselines across extensive sim-to-sim ablations on the Go2 quadruped. Real-world experiments demonstrate substantial drift reduction using less than five minutes of hardware data, even for a challenging two-legged walking behavior. These results demonstrate that proprioceptive distribution matching provides a practical and effective route to simulator adaptation for sim-to-real transfer of learned legged locomotion.

Pranay Dugar, Aayam Shrestha, Fangzhou Yu et al.

The foundational capabilities of humanoid robots should include robustly standing, walking, and mimicry of whole and partial-body motions. This work introduces the Masked Humanoid Controller (MHC), which supports all of these capabilities by tracking target trajectories over selected subsets of humanoid state variables while ensuring balance and robustness against disturbances. The MHC is trained in simulation using a carefully designed curriculum that imitates partially masked motions from a library of behaviors spanning standing, walking, optimized reference trajectories, re-targeted video clips, and human motion capture data. It also allows for combining joystick-based control with partial-body motion mimicry. We showcase simulation experiments validating the MHC's ability to execute a wide variety of behaviors from partially-specified target motions. Moreover, we demonstrate sim-to-real transfer on the real-world Digit V3 humanoid robot. To our knowledge, this is the first instance of a learned controller that can realize whole-body control of a real-world humanoid for such diverse multi-modal targets.

Yassine Chemingui, Aryan Deshwal, Honghao Wei et al.

Offline safe reinforcement learning (OSRL) involves learning a decision-making policy to maximize rewards from a fixed batch of training data to satisfy pre-defined safety constraints. However, adapting to varying safety constraints during deployment without retraining remains an under-explored challenge. To address this challenge, we introduce constraint-adaptive policy switching (CAPS), a wrapper framework around existing offline RL algorithms. During training, CAPS uses offline data to learn multiple policies with a shared representation that optimize different reward and cost trade-offs. During testing, CAPS switches between those policies by selecting at each state the policy that maximizes future rewards among those that satisfy the current cost constraint. Our experiments on 38 tasks from the DSRL benchmark demonstrate that CAPS consistently outperforms existing methods, establishing a strong wrapper-based baseline for OSRL. The code is publicly available at https://github.com/yassineCh/CAPS.

Krishu K. Thapa, Bhupinderjeet Singh, Supriya Savalkar et al.

Snow Water-Equivalent (SWE) -- the amount of water available if snowpack is melted -- is a key decision variable used by water management agencies to make irrigation, flood control, power generation and drought management decisions. SWE values vary spatiotemporally -- affected by weather, topography and other environmental factors. While daily SWE can be measured by Snow Telemetry (SNOTEL) stations with requisite instrumentation, such stations are spatially sparse requiring interpolation techniques to create spatiotemporally complete data. While recent efforts have explored machine learning (ML) for SWE prediction, a number of recent ML advances have yet to be considered. The main contribution of this paper is to explore one such ML advance, attention mechanisms, for SWE prediction. Our hypothesis is that attention has a unique ability to capture and exploit correlations that may exist across locations or the temporal spectrum (or both). We present a generic attention-based modeling framework for SWE prediction and adapt it to capture spatial attention and temporal attention. Our experimental results on 323 SNOTEL stations in the Western U.S. demonstrate that our attention-based models outperform other machine learning approaches. We also provide key results highlighting the differences between spatial and temporal attention in this context and a roadmap toward deployment for generating spatially-complete SWE maps.

Ashish Malik, Caleb Lowe, Aayam Shrestha et al.

We study long-horizon planning in 3D environments from under-specified natural-language goals using only visual observations, focusing on multi-step 3D box rearrangement tasks. Existing approaches typically rely on symbolic planners with brittle relational grounding of states and goals, or on direct action-sequence generation from 2D vision-language models (VLMs). Both approaches struggle with reasoning over many objects, rich 3D geometry, and implicit semantic constraints. Recent advances in 3D VLMs demonstrate strong grounding of natural-language referents to 3D segmentation masks, suggesting the potential for more general planning capabilities. We extend existing 3D grounding models and propose Reactive Action Mask Planner (RAMP-3D), which formulates long-horizon planning as sequential reactive prediction of paired 3D masks: a "which-object" mask indicating what to pick and a "which-target-region" mask specifying where to place it. The resulting system processes RGB-D observations and natural-language task specifications to reactively generate multi-step pick-and-place actions for 3D box rearrangement. We conduct experiments across 11 task variants in warehouse-style environments with 1-30 boxes and diverse natural-language constraints. RAMP-3D achieves 79.5% success rate on long-horizon rearrangement tasks and significantly outperforms 2D VLM-based baselines, establishing mask-based reactive policies as a promising alternative to symbolic pipelines for long-horizon planning.

Yassine Chemingui, Aryan Deshwal, Alan Fern et al.

We study the problem of Offline Safe Reinforcement Learning (OSRL), where the goal is to learn a reward-maximizing policy from fixed data under a cumulative cost constraint. We propose a novel OSRL approach that frames the problem as a minimax objective and solves it by combining offline RL with online optimization algorithms. We prove the approximate optimality of this approach when integrated with an approximate offline RL oracle and no-regret online optimization. We also present a practical approximation that can be combined with any offline RL algorithm, eliminating the need for offline policy evaluation. Empirical results on the DSRL benchmark demonstrate that our method reliably enforces safety constraints under stringent cost budgets, while achieving high rewards. The code is available at https://github.com/yassineCh/O3SRL.

Devin Crowley, Jeremy Dao, Helei Duan et al.

In this paper, we explore the space of running gaits for the bipedal robot Cassie. Our first contribution is to present an approach for optimizing gait efficiency across a spectrum of speeds with the aim of enabling extremely high-speed running on hardware. This raises the question of how the resulting gaits compare to human running mechanics, which are known to be highly efficient in comparison to quadrupeds. Our second contribution is to conduct this comparison based on established human biomechanical studies. We find that despite morphological differences between Cassie and humans, key properties of the gaits are highly similar across a wide range of speeds. Finally, our third contribution is to integrate the optimized running gaits into a full controller that satisfies the rules of the real-world task of the 100m dash, including starting and stopping from a standing position. We demonstrate this controller on hardware to establish the Guinness World Record for Fastest 100m by a Bipedal Robot.

Aayam Shrestha, Pan Liu, German Ros et al.

This work focuses on generating realistic, physically-based human behaviors from multi-modal inputs, which may only partially specify the desired motion. For example, the input may come from a VR controller providing arm motion and body velocity, partial key-point animation, computer vision applied to videos, or even higher-level motion goals. This requires a versatile low-level humanoid controller that can handle such sparse, under-specified guidance, seamlessly switch between skills, and recover from failures. Current approaches for learning humanoid controllers from demonstration data capture some of these characteristics, but none achieve them all. To this end, we introduce the Masked Humanoid Controller (MHC), a novel approach that applies multi-objective imitation learning on augmented and selectively masked motion demonstrations. The training methodology results in an MHC that exhibits the key capabilities of catch-up to out-of-sync input commands, combining elements from multiple motion sequences, and completing unspecified parts of motions from sparse multimodal input. We demonstrate these key capabilities for an MHC learned over a dataset of 87 diverse skills and showcase different multi-modal use cases, including integration with planning frameworks to highlight MHC's ability to solve new user-defined tasks without any finetuning.

Mohammad Rafid Ul Islam, Prasad Tadepalli, Alan Fern

Missing values in multivariate time series data can harm machine learning performance and introduce bias. These gaps arise from sensor malfunctions, blackouts, and human error and are typically addressed by data imputation. Previous work has tackled the imputation of missing data in random, complete blackouts and forecasting scenarios. The current paper addresses a more general missing pattern, which we call "partial blackout," where a subset of features is missing for consecutive time steps. We introduce a two-stage imputation process using self-attention and diffusion processes to model feature and temporal correlations. Notably, our model effectively handles missing data during training, enhancing adaptability and ensuring reliable imputation and performance, even with incomplete datasets. Our experiments on benchmark and two real-world time series datasets demonstrate that our model outperforms the state-of-the-art in partial blackout scenarios and shows better scalability.

William Solow, Sandhya Saisubramanian, Alan Fern

We introduce WOFOSTGym, a novel crop simulation environment designed to train reinforcement learning (RL) agents to optimize agromanagement decisions for annual and perennial crops in single and multi-farm settings. Effective crop management requires optimizing yield and economic returns while minimizing environmental impact, a complex sequential decision-making problem well suited for RL. However, the lack of simulators for perennial crops in multi-farm contexts has hindered RL applications in this domain. Existing crop simulators also do not support multiple annual crops. WOFOSTGym addresses these gaps by supporting 23 annual crops and two perennial crops, enabling RL agents to learn diverse agromanagement strategies in multi-year, multi-crop, and multi-farm settings. Our simulator offers a suite of challenging tasks for learning under partial observability, non-Markovian dynamics, and delayed feedback. WOFOSTGym's standard RL interface allows researchers without agricultural expertise to explore a wide range of agromanagement problems. Our experiments demonstrate the learned behaviors across various crop varieties and soil types, highlighting WOFOSTGym's potential for advancing RL-driven decision support in agriculture.

Mohitvishnu S. Gadde, Pranay Dugar, Ashish Malik et al.

Effective bipedal locomotion in dynamic environments, such as cluttered indoor spaces or uneven terrain, requires agile and adaptive movement in all directions. This necessitates omnidirectional terrain sensing and a controller capable of processing such input. We present a learning framework for vision-based omnidirectional bipedal locomotion, enabling seamless movement using depth images. A key challenge is the high computational cost of rendering omnidirectional depth images in simulation, making traditional sim-to-real reinforcement learning (RL) impractical. Our method combines a robust blind controller with a teacher policy that supervises a vision-based student policy, trained on noise-augmented terrain data to avoid rendering costs during RL and ensure robustness. We also introduce a data augmentation technique for supervised student training, accelerating training by up to 10 times compared to conventional methods. Our framework is validated through simulation and real-world tests, demonstrating effective omnidirectional locomotion with minimal reliance on expensive rendering. This is, to the best of our knowledge, the first demonstration of vision-based omnidirectional bipedal locomotion, showcasing its adaptability to diverse terrains.

Rajesh Mangannavar, Stefan Lee, Alan Fern et al.

We propose a novel approach to learn relational policies for classical planning based on learning to rank actions. We introduce a new graph representation that explicitly captures action information and propose a Graph Neural Network (GNN) architecture augmented with Gated Recurrent Units (GRUs) to learn action rankings. Unlike value-function based approaches that must learn a globally consistent function, our action ranking method only needs to learn locally consistent ranking. Our model is trained on data generated from small problem instances that are easily solved by planners and is applied to significantly larger instances where planning is computationally prohibitive. Experimental results across standard planning benchmarks demonstrate that our action-ranking approach not only achieves better generalization to larger problems than those used in training but also outperforms multiple baselines (value function and action ranking) methods in terms of success rate and plan quality.

Rajesh Mangannavar, Alan Fern, Prasad Tadepalli

We present an online planning framework and a new benchmark dataset for solving multi-object rearrangement problems in partially observable, multi-room environments. Current object rearrangement solutions, primarily based on Reinforcement Learning or hand-coded planning methods, often lack adaptability to diverse challenges. To address this limitation, we introduce a novel Hierarchical Object-Oriented Partially Observed Markov Decision Process (HOO-POMDP) planning approach. This approach comprises of (a) an object-oriented POMDP planner generating sub-goals, (b) a set of low-level policies for sub-goal achievement, and (c) an abstraction system converting the continuous low-level world into a representation suitable for abstract planning. To enable rigorous evaluation of rearrangement challenges, we introduce MultiRoomR, a comprehensive benchmark featuring diverse multi-room environments with varying degrees of partial observability (10-30\% initial visibility), blocked paths, obstructed goals, and multiple objects (10-20) distributed across 2-4 rooms. Experiments demonstrate that our system effectively handles these complex scenarios while maintaining robust performance even with imperfect perception, achieving promising results across both existing benchmarks and our new MultiRoomR dataset.

Pranay Dugar, Mohitvishnu S. Gadde, Jonah Siekmann et al.

Humanoids operating in real-world workspaces must frequently execute task-driven, short-range movements to SE(2) target poses. To be practical, these transitions must be fast, robust, and energy efficient. While learning-based locomotion has made significant progress, most existing methods optimize for velocity-tracking rather than direct pose reaching, resulting in inefficient, marching-style behavior when applied to short-range tasks. In this work, we develop a reinforcement learning approach that directly optimizes humanoid locomotion for SE(2) targets. Central to this approach is a new constellation-based reward function that encourages natural and efficient target-oriented movement. To evaluate performance, we introduce a benchmarking framework that measures energy consumption, time-to-target, and footstep count on a distribution of SE(2) goals. Our results show that the proposed approach consistently outperforms standard methods and enables successful transfer from simulation to hardware, highlighting the importance of targeted reward design for practical short-range humanoid locomotion.

Kristen Goebel, William Solow, Paola Pesantez-Cabrera et al.

This paper introduces a novel approach to budgeted online active learning from finite-horizon data streams with extremely limited labeling budgets. In agricultural applications, such streams might include daily weather data over a growing season, and labels require costly measurements of weather-dependent plant characteristics. Our method integrates two key sources of prior information: a collection of preexisting expert predictors and episodic behavioral knowledge of the experts based on unlabeled data streams. Unlike previous research on online active learning with experts, our work simultaneously considers query budgets, finite horizons, and episodic knowledge, enabling effective learning in applications with severely limited labeling capacity. We demonstrate the utility of our approach through experiments on various prediction problems derived from both a realistic agricultural crop simulator and real-world data from multiple grape cultivars. The results show that our method significantly outperforms baseline expert predictions, uniform query selection, and existing approaches that consider budgets and limited horizons but neglect episodic knowledge, even under highly constrained labeling budgets.

Kristen Goebel, Paola Pesantez-Cabrera, Markus Keller et al.

Cold temperatures can cause significant frost damage to fruit crops depending on their resilience, or cold hardiness, which changes throughout the dormancy season. This has led to the development of predictive cold-hardiness models, which help farmers decide when to deploy expensive frost-mitigation measures. Unfortunately, cold-hardiness data for model training is only available for some fruit cultivars due to the need for specialized equipment and expertise. Rather, farmers often do have years of phenological data (e.g. date of budbreak) that they regularly collect for their crops. In this work, we introduce a new transfer-learning framework, Transfer via Auxiliary Labels (TAL), that allows farmers to leverage the phenological data to produce more accurate cold-hardiness predictions, even when no cold-hardiness data is available for their specific crop. The framework assumes a set of source tasks (cultivars) where each has associated primary labels (cold hardiness) and auxiliary labels (phenology). However, the target task (new cultivar) is assumed to only have the auxiliary labels. The goal of TAL is to predict primary labels for the target task via transfer from the source tasks. Surprisingly, despite the vast literature on transfer learning, to our knowledge, the TAL formulation has not been previously addressed. Thus, we propose several new TAL approaches based on model selection and averaging that can leverage recent deep multi-task models for cold-hardiness prediction. Our results on real-world cold-hardiness and phenological data for multiple grape cultivars demonstrate that TAL can leverage the phenological data to improve cold-hardiness predictions in the absence of cold-hardiness data.

Bikram Pandit, Ashutosh Gupta, Mohitvishnu S. Gadde et al.

Payload transport over flat terrain via multi-wheel robot carriers is well-understood, highly effective, and configurable. In this paper, our goal is to provide similar effectiveness and configurability for transport over rough terrain that is more suitable for legs rather than wheels. For this purpose, we consider multi-biped robot carriers, where wheels are replaced by multiple bipedal robots attached to the carrier. Our main contribution is to design a decentralized controller for such systems that can be effectively applied to varying numbers and configurations of rigidly attached bipedal robots without retraining. We present a reinforcement learning approach for training the controller in simulation that supports transfer to the real world. Our experiments in simulation provide quantitative metrics showing the effectiveness of the approach over a wide variety of simulated transport scenarios. In addition, we demonstrate the controller in the real-world for systems composed of two and three Cassie robots. To our knowledge, this is the first example of a scalable multi-biped payload transport system.

Kin-Ho Lam, Zhengxian Lin, Jed Irvine et al.

Enabling humans to identify potential flaws in an agent's decision making is an important Explainable AI application. We consider identifying such flaws in a planning-based deep reinforcement learning (RL) agent for a complex real-time strategy game. In particular, the agent makes decisions via tree search using a learned model and evaluation function over interpretable states and actions. This gives the potential for humans to identify flaws at the level of reasoning steps in the tree, even if the entire reasoning process is too complex to understand. However, it is unclear whether humans will be able to identify such flaws due to the size and complexity of trees. We describe a user interface and case study, where a small group of AI experts and developers attempt to identify reasoning flaws due to inaccurate agent learning. Overall, the interface allowed the group to identify a number of significant flaws of varying types, demonstrating the promise of this approach.

Li Fuxin, Zhongang Qi, Saeed Khorram et al.

This paper summarizes our endeavors in the past few years in terms of explaining image classifiers, with the aim of including negative results and insights we have gained. The paper starts with describing the explainable neural network (XNN), which attempts to extract and visualize several high-level concepts purely from the deep network, without relying on human linguistic concepts. This helps users understand network classifications that are less intuitive and substantially improves user performance on a difficult fine-grained classification task of discriminating among different species of seagulls. Realizing that an important missing piece is a reliable heatmap visualization tool, we have developed I-GOS and iGOS++ utilizing integrated gradients to avoid local optima in heatmap generation, which improved the performance across all resolutions. During the development of those visualizations, we realized that for a significant number of images, the classifier has multiple different paths to reach a confident prediction. This has lead to our recent development of structured attention graphs (SAGs), an approach that utilizes beam search to locate multiple coarse heatmaps for a single image, and compactly visualizes a set of heatmaps by capturing how different combinations of image regions impact the confidence of a classifier. Through the research process, we have learned much about insights in building deep network explanations, the existence and frequency of multiple explanations, and various tricks of the trade that make explanations work. In this paper, we attempt to share those insights and opinions with the readers with the hope that some of them will be informative for future researchers on explainable deep learning.

Mohamad H Danesh, Alan Fern

We study the problem of out-of-distribution dynamics (OODD) detection, which involves detecting when the dynamics of a temporal process change compared to the training-distribution dynamics. This is relevant to applications in control, reinforcement learning (RL), and multi-variate time-series, where changes to test time dynamics can impact the performance of learning controllers/predictors in unknown ways. This problem is particularly important in the context of deep RL, where learned controllers often overfit to the training environment. Currently, however, there is a lack of established OODD benchmarks for the types of environments commonly used in RL research. Our first contribution is to design a set of OODD benchmarks derived from common RL environments with varying types and intensities of OODD. Our second contribution is to design a strong OODD baseline approach based on recurrent implicit quantile network (RIQN), which monitors autoregressive prediction errors for OODD detection. In addition to RIQN, we introduce and test three other simpler baselines. Our final contribution is to evaluate our baseline approaches on the benchmarks to provide results for future comparison.

Sam Greydanus, Stefan Lee, Alan Fern

Neural networks are a popular tool for modeling sequential data but they generally do not treat time as a continuous variable. Neural ODEs represent an important exception: they parameterize the time derivative of a hidden state with a neural network and then integrate over arbitrary amounts of time. But these parameterizations, which have arbitrary curvature, can be hard to integrate and thus train and evaluate. In this paper, we propose making a piecewise-constant approximation to Neural ODEs to mitigate these issues. Our model can be integrated exactly via Euler integration and can generate autoregressive samples in 3-20 times fewer steps than comparable RNN and ODE-RNN models. We evaluate our model on several synthetic physics tasks and a planning task inspired by the game of billiards. We find that it matches the performance of baseline approaches while requiring less time to train and evaluate.

Jonah Siekmann, Kevin Green, John Warila et al.

Accurate and precise terrain estimation is a difficult problem for robot locomotion in real-world environments. Thus, it is useful to have systems that do not depend on accurate estimation to the point of fragility. In this paper, we explore the limits of such an approach by investigating the problem of traversing stair-like terrain without any external perception or terrain models on a bipedal robot. For such blind bipedal platforms, the problem appears difficult (even for humans) due to the surprise elevation changes. Our main contribution is to show that sim-to-real reinforcement learning (RL) can achieve robust locomotion over stair-like terrain on the bipedal robot Cassie using only proprioceptive feedback. Importantly, this only requires modifying an existing flat-terrain training RL framework to include stair-like terrain randomization, without any changes in reward function. To our knowledge, this is the first controller for a bipedal, human-scale robot capable of reliably traversing a variety of real-world stairs and other stair-like disturbances using only proprioception.

Erich Merrill, Stefan Lee, Li Fuxin et al.

We consider the problem of modeling the dynamics of continuous spatial-temporal processes represented by irregular samples through both space and time. Such processes occur in sensor networks, citizen science, multi-robot systems, and many others. We propose a new deep model that is able to directly learn and predict over this irregularly sampled data, without voxelization, by leveraging a recent convolutional architecture for static point clouds. The model also easily incorporates the notion of multiple entities in the process. In particular, the model can flexibly answer prediction queries about arbitrary space-time points for different entities regardless of the distribution of the training or test-time data. We present experiments on real-world weather station data and battles between large armies in StarCraft II. The results demonstrate the model's flexibility in answering a variety of query types and demonstrate improved performance and efficiency compared to state-of-the-art baselines.

Aryan Deshwal, Syrine Belakaria, Janardhan Rao Doppa et al.

We consider the problem of optimizing expensive black-box functions over discrete spaces (e.g., sets, sequences, graphs). The key challenge is to select a sequence of combinatorial structures to evaluate, in order to identify high-performing structures as quickly as possible. Our main contribution is to introduce and evaluate a new learning-to-search framework for this problem called L2S-DISCO. The key insight is to employ search procedures guided by control knowledge at each step to select the next structure and to improve the control knowledge as new function evaluations are observed. We provide a concrete instantiation of L2S-DISCO for local search procedure and empirically evaluate it on diverse real-world benchmarks. Results show the efficacy of L2S-DISCO over state-of-the-art algorithms in solving complex optimization problems.

Vivswan Shitole, Li Fuxin, Minsuk Kahng et al.

Attention maps are a popular way of explaining the decisions of convolutional networks for image classification. Typically, for each image of interest, a single attention map is produced, which assigns weights to pixels based on their importance to the classification. A single attention map, however, provides an incomplete understanding since there are often many other maps that explain a classification equally well. In this paper, we introduce structured attention graphs (SAGs), which compactly represent sets of attention maps for an image by capturing how different combinations of image regions impact a classifier's confidence. We propose an approach to compute SAGs and a visualization for SAGs so that deeper insight can be gained into a classifier's decisions. We conduct a user study comparing the use of SAGs to traditional attention maps for answering counterfactual questions about image classifications. Our results show that the users are more correct when answering comparative counterfactual questions based on SAGs compared to the baselines.

Helei Duan, Jeremy Dao, Kevin Green et al.

Recent work has demonstrated the success of reinforcement learning (RL) for training bipedal locomotion policies for real robots. This prior work, however, has focused on learning joint-coordination controllers based on an objective of following joint trajectories produced by already available controllers. As such, it is difficult to train these approaches to achieve higher-level goals of legged locomotion, such as simply specifying the desired end-effector foot movement or ground reaction forces. In this work, we propose an approach for integrating knowledge of the robot system into RL to allow for learning at the level of task space actions in terms of feet setpoints. In particular, we integrate learning a task space policy with a model-based inverse dynamics controller, which translates task space actions into joint-level controls. With this natural action space for learning locomotion, the approach is more sample efficient and produces desired task space dynamics compared to learning purely joint space actions. We demonstrate the approach in simulation and also show that the learned policies are able to transfer to the real bipedal robot Cassie. This result encourages further research towards incorporating bipedal control techniques into the structure of the learning process to enable dynamic behaviors.

Jonah Siekmann, Yesh Godse, Alan Fern et al.

We study the problem of realizing the full spectrum of bipedal locomotion on a real robot with sim-to-real reinforcement learning (RL). A key challenge of learning legged locomotion is describing different gaits, via reward functions, in a way that is intuitive for the designer and specific enough to reliably learn the gait across different initial random seeds or hyperparameters. A common approach is to use reference motions (e.g. trajectories of joint positions) to guide learning. However, finding high-quality reference motions can be difficult and the trajectories themselves narrowly constrain the space of learned motion. At the other extreme, reference-free reward functions are often underspecified (e.g. move forward) leading to massive variance in policy behavior, or are the product of significant reward-shaping via trial-and-error, making them exclusive to specific gaits. In this work, we propose a reward-specification framework based on composing simple probabilistic periodic costs on basic forces and velocities. We instantiate this framework to define a parametric reward function with intuitive settings for all common bipedal gaits - standing, walking, hopping, running, and skipping. Using this function we demonstrate successful sim-to-real transfer of the learned gaits to the bipedal robot Cassie, as well as a generic policy that can transition between all of the two-beat gaits.

Kevin Green, Yesh Godse, Jeremy Dao et al.

In this paper, we describe an approach to achieve dynamic legged locomotion on physical robots which combines existing methods for control with reinforcement learning. Specifically, our goal is a control hierarchy in which highest-level behaviors are planned through reduced-order models, which describe the fundamental physics of legged locomotion, and lower level controllers utilize a learned policy that can bridge the gap between the idealized, simple model and the complex, full order robot. The high-level planner can use a model of the environment and be task specific, while the low-level learned controller can execute a wide range of motions so that it applies to many different tasks. In this letter we describe this learned dynamic walking controller and show that a range of walking motions from reduced-order models can be used as the command and primary training signal for learned policies. The resulting policies do not attempt to naively track the motion (as a traditional trajectory tracking controller would) but instead balance immediate motion tracking with long term stability. The resulting controller is demonstrated on a human scale, unconstrained, untethered bipedal robot at speeds up to 1.2 m/s. This letter builds the foundation of a generic, dynamic learned walking controller that can be applied to many different tasks.

Anurag Koul, Varun V. Kumar, Alan Fern et al.

Learning and planning with latent space dynamics has been shown to be useful for sample efficiency in model-based reinforcement learning (MBRL) for discrete and continuous control tasks. In particular, recent work, for discrete action spaces, demonstrated the effectiveness of latent-space planning via Monte-Carlo Tree Search (MCTS) for bootstrapping MBRL during learning and at test time. However, the potential gains from latent-space tree search have not yet been demonstrated for environments with continuous action spaces. In this work, we propose and explore an MBRL approach for continuous action spaces based on tree-based planning over learned latent dynamics. We show that it is possible to demonstrate the types of bootstrapping benefits as previously shown for discrete spaces. In particular, the approach achieves improved sample efficiency and performance on a majority of challenging continuous-control benchmarks compared to the state-of-the-art.

Aayam Shrestha, Stefan Lee, Prasad Tadepalli et al.

We study an approach to offline reinforcement learning (RL) based on optimally solving finitely-represented MDPs derived from a static dataset of experience. This approach can be applied on top of any learned representation and has the potential to easily support multiple solution objectives as well as zero-shot adjustment to changing environments and goals. Our main contribution is to introduce the Deep Averagers with Costs MDP (DAC-MDP) and to investigate its solutions for offline RL. DAC-MDPs are a non-parametric model that can leverage deep representations and account for limited data by introducing costs for exploiting under-represented parts of the model. In theory, we show conditions that allow for lower-bounding the performance of DAC-MDP solutions. We also investigate the empirical behavior in a number of environments, including those with image-based observations. Overall, the experiments demonstrate that the framework can work in practice and scale to large complex offline RL problems.

Zhengxian Lin, Kim-Ho Lam, Alan Fern

We investigate a deep reinforcement learning (RL) architecture that supports explaining why a learned agent prefers one action over another. The key idea is to learn action-values that are directly represented via human-understandable properties of expected futures. This is realized via the embedded self-prediction (ESP)model, which learns said properties in terms of human provided features. Action preferences can then be explained by contrasting the future properties predicted for each action. To address cases where there are a large number of features, we develop a novel method for computing minimal sufficient explanations from anESP. Our case studies in three domains, including a complex strategy game, show that ESP models can be effectively learned and support insightful explanations.

Mohamad H. Danesh, Anurag Koul, Alan Fern et al.

We introduce an approach for understanding control policies represented as recurrent neural networks. Recent work has approached this problem by transforming such recurrent policy networks into finite-state machines (FSM) and then analyzing the equivalent minimized FSM. While this led to interesting insights, the minimization process can obscure a deeper understanding of a machine's operation by merging states that are semantically distinct. To address this issue, we introduce an analysis approach that starts with an unminimized FSM and applies more-interpretable reductions that preserve the key decision points of the policy. We also contribute an attention tool to attain a deeper understanding of the role of observations in the decisions. Our case studies on 7 Atari games and 3 control benchmarks demonstrate that the approach can reveal insights that have not been previously noticed.

Jonah Siekmann, Srikar Valluri, Jeremy Dao et al.

Controlling a non-statically stable biped is a difficult problem largely due to the complex hybrid dynamics involved. Recent work has demonstrated the effectiveness of reinforcement learning (RL) for simulation-based training of neural network controllers that successfully transfer to real bipeds. The existing work, however, has primarily used simple memoryless network architectures, even though more sophisticated architectures, such as those including memory, often yield superior performance in other RL domains. In this work, we consider recurrent neural networks (RNNs) for sim-to-real biped locomotion, allowing for policies that learn to use internal memory to model important physical properties. We show that while RNNs are able to significantly outperform memoryless policies in simulation, they do not exhibit superior behavior on the real biped due to overfitting to the simulation physics unless trained using dynamics randomization to prevent overfitting; this leads to consistently better sim-to-real transfer. We also show that RNNs could use their learned memory states to perform online system identification by encoding parameters of the dynamics into memory.

Murugeswari Issakkimuthu, Alan Fern, Prasad Tadepalli

There are notable examples of online search improving over hand-coded or learned policies (e.g. AlphaZero) for sequential decision making. It is not clear, however, whether or not policy improvement is guaranteed for many of these approaches, even when given a perfect evaluation function and transition model. Indeed, simple counter examples show that seemingly reasonable online search procedures can hurt performance compared to the original policy. To address this issue, we introduce the choice function framework for analyzing online search procedures for policy improvement. A choice function specifies the actions to be considered at every node of a search tree, with all other actions being pruned. Our main contribution is to give sufficient conditions for stationary and non-stationary choice functions to guarantee that the value achieved by online search is no worse than the original policy. In addition, we describe a general parametric class of choice functions that satisfy those conditions and present an illustrative use case of the framework's empirical utility.

Andrew Anderson, Jonathan Dodge, Amrita Sadarangani et al.

We present a user study to investigate the impact of explanations on non-experts' understanding of reinforcement learning (RL) agents. We investigate both a common RL visualization, saliency maps (the focus of attention), and a more recent explanation type, reward-decomposition bars (predictions of future types of rewards). We designed a 124 participant, four-treatment experiment to compare participants' mental models of an RL agent in a simple Real-Time Strategy (RTS) game. Our results show that the combination of both saliency and reward bars were needed to achieve a statistically significant improvement in mental model score over the control. In addition, our qualitative analysis of the data reveals a number of effects for further study.

Mandana Hamidi-Haines, Zhongang Qi, Alan Fern et al.

We consider the problem of explaining the decisions of deep neural networks for image recognition in terms of human-recognizable visual concepts. In particular, given a test set of images, we aim to explain each classification in terms of a small number of image regions, or activation maps, which have been associated with semantic concepts by a human annotator. This allows for generating summary views of the typical reasons for classifications, which can help build trust in a classifier and/or identify example types for which the classifier may not be trusted. For this purpose, we developed a user interface for "interactive naming," which allows a human annotator to manually cluster significant activation maps in a test set into meaningful groups called "visual concepts". The main contribution of this paper is a systematic study of the visual concepts produced by five human annotators using the interactive naming interface. In particular, we consider the adequacy of the concepts for explaining the classification of test-set images, correspondence of the concepts to activations of individual neurons, and the inter-annotator agreement of visual concepts. We find that a large fraction of the activation maps have recognizable visual concepts, and that there is significant agreement between the different annotators about their denotations. Our work is an exploratory study of the interplay between machine learning and human recognition mediated by visualizations of the results of learning.

Anurag Koul, Sam Greydanus, Alan Fern

Recurrent neural networks (RNNs) are an effective representation of control policies for a wide range of reinforcement and imitation learning problems. RNN policies, however, are particularly difficult to explain, understand, and analyze due to their use of continuous-valued memory vectors and observation features. In this paper, we introduce a new technique, Quantized Bottleneck Insertion, to learn finite representations of these vectors and features. The result is a quantized representation of the RNN that can be analyzed to improve our understanding of memory use and general behavior. We present results of this approach on synthetic environments and six Atari games. The resulting finite representations are surprisingly small in some cases, using as few as 3 discrete memory states and 10 observations for a perfect Pong policy. We also show that these finite policy representations lead to improved interpretability.

Si Liu, Risheek Garrepalli, Thomas G. Dietterich et al.

Open category detection is the problem of detecting "alien" test instances that belong to categories or classes that were not present in the training data. In many applications, reliably detecting such aliens is central to ensuring the safety and accuracy of test set predictions. Unfortunately, there are no algorithms that provide theoretical guarantees on their ability to detect aliens under general assumptions. Further, while there are algorithms for open category detection, there are few empirical results that directly report alien detection rates. Thus, there are significant theoretical and empirical gaps in our understanding of open category detection. In this paper, we take a step toward addressing this gap by studying a simple, but practically-relevant variant of open category detection. In our setting, we are provided with a "clean" training set that contains only the target categories of interest and an unlabeled "contaminated" training set that contains a fraction $α$ of alien examples. Under the assumption that we know an upper bound on $α$, we develop an algorithm with PAC-style guarantees on the alien detection rate, while aiming to minimize false alarms. Empirical results on synthetic and standard benchmark datasets demonstrate the regimes in which the algorithm can be effective and provide a baseline for further advancements.

Sam Greydanus, Anurag Koul, Jonathan Dodge et al.

While deep reinforcement learning (deep RL) agents are effective at maximizing rewards, it is often unclear what strategies they use to do so. In this paper, we take a step toward explaining deep RL agents through a case study using Atari 2600 environments. In particular, we focus on using saliency maps to understand how an agent learns and executes a policy. We introduce a method for generating useful saliency maps and use it to show 1) what strong agents attend to, 2) whether agents are making decisions for the right or wrong reasons, and 3) how agents evolve during learning. We also test our method on non-expert human subjects and find that it improves their ability to reason about these agents. Overall, our results show that saliency information can provide significant insight into an RL agent's decisions and learning behavior.

Jose Picado, Arash Termehchy, Sudhanshu Pathak et al.

Relational databases are valuable resources for learning novel and interesting relations and concepts. In order to constraint the search through the large space of candidate definitions, users must tune the algorithm by specifying a language bias. Unfortunately, specifying the language bias is done via trial and error and is guided by the expert's intuitions. We propose AutoBias, a system that leverages information in the schema and content of the database to automatically induce the language bias used by popular relational learning systems. We show that AutoBias delivers the same accuracy as using manually-written language bias by imposing only a slight overhead on the running time of the learning algorithm.

Shubhomoy Das, Weng-Keen Wong, Alan Fern et al.

Anomaly detectors are often used to produce a ranked list of statistical anomalies, which are examined by human analysts in order to extract the actual anomalies of interest. Unfortunately, in realworld applications, this process can be exceedingly difficult for the analyst since a large fraction of high-ranking anomalies are false positives and not interesting from the application perspective. In this paper, we aim to make the analyst's job easier by allowing for analyst feedback during the investigation process. Ideally, the feedback influences the ranking of the anomaly detector in a way that reduces the number of false positives that must be examined before discovering the anomalies of interest. In particular, we introduce a novel technique for incorporating simple binary feedback into tree-based anomaly detectors. We focus on the Isolation Forest algorithm as a representative tree-based anomaly detector, and show that we can significantly improve its performance by incorporating feedback, when compared with the baseline algorithm that does not incorporate feedback. Our technique is simple and scales well as the size of the data increases, which makes it suitable for interactive discovery of anomalies in large datasets.

Behrooz Mahasseni, Sinisa Todorovic, Alan Fern

This paper formulates and presents a solution to the new problem of budgeted semantic video segmentation. Given a video, the goal is to accurately assign a semantic class label to every pixel in the video within a specified time budget. Typical approaches to such labeling problems, such as Conditional Random Fields (CRFs), focus on maximizing accuracy but do not provide a principled method for satisfying a time budget. For video data, the time required by CRF and related methods is often dominated by the time to compute low-level descriptors of supervoxels across the video. Our key contribution is the new budgeted inference framework for CRF models that intelligently selects the most useful subsets of descriptors to run on subsets of supervoxels within the time budget. The objective is to maintain an accuracy as close as possible to the CRF model with no time bound, while remaining within the time budget. Our second contribution is the algorithm for learning a policy for the sparse selection of supervoxels and their descriptors for budgeted CRF inference. This learning algorithm is derived by casting our problem in the framework of Markov Decision Processes, and then instantiating a state-of-the-art policy learning algorithm known as Classification-Based Approximate Policy Iteration. Our experiments on multiple video datasets show that our learning approach and framework is able to significantly reduce computation time, and maintain competitive accuracy under varying budgets.

Jose Picado, Arash Termehchy, Alan Fern et al.

Learning novel concepts and relations from relational databases is an important problem with many applications in database systems and machine learning. Relational learning algorithms learn the definition of a new relation in terms of existing relations in the database. Nevertheless, the same data set may be represented under different schemas for various reasons, such as efficiency, data quality, and usability. Unfortunately, the output of current relational learning algorithms tends to vary quite substantially over the choice of schema, both in terms of learning accuracy and efficiency. This variation complicates their off-the-shelf application. In this paper, we introduce and formalize the property of schema independence of relational learning algorithms, and study both the theoretical and empirical dependence of existing algorithms on the common class of (de) composition schema transformations. We study both sample-based learning algorithms, which learn from sets of labeled examples, and query-based algorithms, which learn by asking queries to an oracle. We prove that current relational learning algorithms are generally not schema independent. For query-based learning algorithms we show that the (de) composition transformations influence their query complexity. We propose Castor, a sample-based relational learning algorithm that achieves schema independence by leveraging data dependencies. We support the theoretical results with an empirical study that demonstrates the schema dependence/independence of several algorithms on existing benchmark and real-world datasets under (de) compositions.