Towards Automated Semantic Interpretability in Reinforcement Learning via Vision-Language Models

Semantic Interpretability in Reinforcement Learning (RL) enables transparency, accountability, and safer deployment by making the agent's decisions understandable and verifiable. Achieving this, however, requires a feature space composed of human-understandable concepts, which traditionally rely on human specification and fail to generalize to unseen environments. In this work, we introduce Semantically Interpretable Reinforcement Learning with Vision-Language Models Empowered Automation (SILVA), an automated framework that leverages pre-trained vision-language models (VLM) for semantic feature extraction and interpretable tree-based models for policy optimization. SILVA first queries a VLM to identify relevant semantic features for an unseen environment, then extracts these features from the environment. Finally, it trains an Interpretable Control Tree via RL, mapping the extracted features to actions in a transparent and interpretable manner. To address the computational inefficiency of extracting features directly with VLMs, we develop a feature extraction pipeline that generates a dataset for training a lightweight convolutional network, which is subsequently used during RL. By leveraging VLMs to automate tree-based RL, SILVA removes the reliance on human annotation previously required by interpretable models while also overcoming the inability of VLMs alone to generate valid robot policies, enabling semantically interpretable reinforcement learning without human-in-the-loop.

ELEMENTAL: Interactive Learning from Demonstrations and Vision-Language Models for Reward Design in Robotics

Reinforcement learning (RL) has demonstrated compelling performance in robotic tasks, but its success often hinges on the design of complex, ad hoc reward functions. Researchers have explored how Large Language Models (LLMs) could enable non-expert users to specify reward functions more easily. However, LLMs struggle to balance the importance of different features, generalize poorly to out-of-distribution robotic tasks, and cannot represent the problem properly with only text-based descriptions. To address these challenges, we propose ELEMENTAL (intEractive LEarning froM dEmoNstraTion And Language), a novel framework that combines natural language guidance with visual user demonstrations to align robot behavior with user intentions better. By incorporating visual inputs, ELEMENTAL overcomes the limitations of text-only task specifications, while leveraging inverse reinforcement learning (IRL) to balance feature weights and match the demonstrated behaviors optimally. ELEMENTAL also introduces an iterative feedback-loop through self-reflection to improve feature, reward, and policy learning. Our experiment results demonstrate that ELEMENTAL outperforms prior work by 42.3% on task success, and achieves 41.3% better generalization in out-of-distribution tasks, highlighting its robustness in LfD.

Faster Model Predictive Control via Self-Supervised Initialization Learning

Optimization for robot control tasks, spanning various methodologies, includes Model Predictive Control (MPC). However, the complexity of the system, such as non-convex and non-differentiable cost functions and prolonged planning horizons often drastically increases the computation time, limiting MPC's real-world applicability. Prior works in speeding up the optimization have limitations on solving convex problem and generalizing to hold out domains. To overcome this challenge, we develop a novel framework aiming at expediting optimization processes. In our framework, we combine offline self-supervised learning and online fine-tuning through reinforcement learning to improve the control performance and reduce optimization time. We demonstrate the effectiveness of our method on a novel, challenging Formula-1-track driving task, achieving 3.9\% higher performance in optimization time and 3.6\% higher performance in tracking accuracy on challenging holdout tracks.

Interpretable Reinforcement Learning for Robotics and Continuous Control

Interpretability in machine learning is critical for the safe deployment of learned policies across legally-regulated and safety-critical domains. While gradient-based approaches in reinforcement learning have achieved tremendous success in learning policies for continuous control problems such as robotics and autonomous driving, the lack of interpretability is a fundamental barrier to adoption. We propose Interpretable Continuous Control Trees (ICCTs), a tree-based model that can be optimized via modern, gradient-based, reinforcement learning approaches to produce high-performing, interpretable policies. The key to our approach is a procedure for allowing direct optimization in a sparse decision-tree-like representation. We validate ICCTs against baselines across six domains, showing that ICCTs are capable of learning policies that parity or outperform baselines by up to 33% in autonomous driving scenarios while achieving a 300x-600x reduction in the number of parameters against deep learning baselines. We prove that ICCTs can serve as universal function approximators and display analytically that ICCTs can be verified in linear time. Furthermore, we deploy ICCTs in two realistic driving domains, based on interstate Highway-94 and 280 in the US. Finally, we verify ICCT's utility with end-users and find that ICCTs are rated easier to simulate, quicker to validate, and more interpretable than neural networks.

Learning Models of Adversarial Agent Behavior under Partial Observability

The need for opponent modeling and tracking arises in several real-world scenarios, such as professional sports, video game design, and drug-trafficking interdiction. In this work, we present Graph based Adversarial Modeling with Mutal Information (GrAMMI) for modeling the behavior of an adversarial opponent agent. GrAMMI is a novel graph neural network (GNN) based approach that uses mutual information maximization as an auxiliary objective to predict the current and future states of an adversarial opponent with partial observability. To evaluate GrAMMI, we design two large-scale, pursuit-evasion domains inspired by real-world scenarios, where a team of heterogeneous agents is tasked with tracking and interdicting a single adversarial agent, and the adversarial agent must evade detection while achieving its own objectives. With the mutual information formulation, GrAMMI outperforms all baselines in both domains and achieves 31.68% higher log-likelihood on average for future adversarial state predictions across both domains.

Adversarial Search and Track with Multiagent Reinforcement Learning in Sparsely Observable Environment

We study a search and tracking (S&T) problem for a team of dynamic search agents to capture an adversarial evasive agent with only sparse temporal and spatial knowledge of its location in this paper. The domain is challenging for traditional Reinforcement Learning (RL) approaches as the large space leads to sparse observations of the adversary and in turn sparse rewards for the search agents. Additionally, the opponent's behavior is reactionary to the search agents, which causes a data distribution shift for RL during training as search agents improve their policies. We propose a differentiable Multi-Agent RL (MARL) architecture that utilizes a novel filtering module to supplement estimated adversary location information and enables the effective learning of a team policy. Our algorithm learns how to balance information from prior knowledge and a motion model to remain resilient to the data distribution shift and outperforms all baseline methods with a 46% increase of detection rate.

The Effect of Robot Skill Level and Communication in Rapid, Proximate Human-Robot Collaboration

As high-speed, agile robots become more commonplace, these robots will have the potential to better aid and collaborate with humans. However, due to the increased agility and functionality of these robots, close collaboration with humans can create safety concerns that alter team dynamics and degrade task performance. In this work, we aim to enable the deployment of safe and trustworthy agile robots that operate in proximity with humans. We do so by 1) Proposing a novel human-robot doubles table tennis scenario to serve as a testbed for studying agile, proximate human-robot collaboration and 2) Conducting a user-study to understand how attributes of the robot (e.g., robot competency or capacity to communicate) impact team dynamics, perceived safety, and perceived trust, and how these latent factors affect human-robot collaboration (HRC) performance. We find that robot competency significantly increases perceived trust ($p<.001$), extending skill-to-trust assessments in prior studies to agile, proximate HRC. Furthermore, interestingly, we find that when the robot vocalizes its intention to perform a task, it results in a significant decrease in team performance ($p=.037$) and perceived safety of the system ($p=.009$).

Utilizing Human Feedback for Primitive Optimization in Wheelchair Tennis

Agile robotics presents a difficult challenge with robots moving at high speeds requiring precise and low-latency sensing and control. Creating agile motion that accomplishes the task at hand while being safe to execute is a key requirement for agile robots to gain human trust. This requires designing new approaches that are flexible and maintain knowledge over world constraints. In this paper, we consider the problem of building a flexible and adaptive controller for a challenging agile mobile manipulation task of hitting ground strokes on a wheelchair tennis robot. We propose and evaluate an extension to work done on learning striking behaviors using a probabilistic movement primitive (ProMP) framework by (1) demonstrating the safe execution of learned primitives on an agile mobile manipulator setup, and (2) proposing an online primitive refinement procedure that utilizes evaluative feedback from humans on the executed trajectories.

Safe Inverse Reinforcement Learning via Control Barrier Function

Learning from Demonstration (LfD) is a powerful method for enabling robots to perform novel tasks as it is often more tractable for a non-roboticist end-user to demonstrate the desired skill and for the robot to efficiently learn from the associated data than for a human to engineer a reward function for the robot to learn the skill via reinforcement learning (RL). Safety issues arise in modern LfD techniques, e.g., Inverse Reinforcement Learning (IRL), just as they do for RL; yet, safe learning in LfD has received little attention. In the context of agile robots, safety is especially vital due to the possibility of robot-environment collision, robot-human collision, and damage to the robot. In this paper, we propose a safe IRL framework, CBFIRL, that leverages the Control Barrier Function (CBF) to enhance the safety of the IRL policy. The core idea of CBFIRL is to combine a loss function inspired by CBF requirements with the objective in an IRL method, both of which are jointly optimized via gradient descent. In the experiments, we show our framework performs safer compared to IRL methods without CBF, that is $\sim15\%$ and $\sim20\%$ improvement for two levels of difficulty of a 2D racecar domain and $\sim 50\%$ improvement for a 3D drone domain.

Athletic Mobile Manipulator System for Robotic Wheelchair Tennis

Athletics are a quintessential and universal expression of humanity. From French monks who in the 12th century invented jeu de paume, the precursor to modern lawn tennis, back to the K'iche' people who played the Maya Ballgame as a form of religious expression over three thousand years ago, humans have sought to train their minds and bodies to excel in sporting contests. Advances in robotics are opening up the possibility of robots in sports. Yet, key challenges remain, as most prior works in robotics for sports are limited to pristine sensing environments, do not require significant force generation, or are on miniaturized scales unsuited for joint human-robot play. In this paper, we propose the first open-source, autonomous robot for playing regulation wheelchair tennis. We demonstrate the performance of our full-stack system in executing ground strokes and evaluate each of the system's hardware and software components. The goal of this paper is to (1) inspire more research in human-scale robot athletics and (2) establish the first baseline towards developing a robot in future work that can serve as a teammate for mixed, human-robot doubles play. Our paper contributes to the science of systems design and poses a set of key challenges for the robotics community to address in striving towards a vision of human-robot collaboration in sports.