SELP: Generating Safe and Efficient Task Plans for Robot Agents with Large Language Models

Despite significant advancements in large language models (LLMs) that enhance robot agents' understanding and execution of natural language (NL) commands, ensuring the agents adhere to user-specified constraints remains challenging, particularly for complex commands and long-horizon tasks. To address this challenge, we present three key insights, equivalence voting, constrained decoding, and domain-specific fine-tuning, which significantly enhance LLM planners' capability in handling complex tasks. Equivalence voting ensures consistency by generating and sampling multiple Linear Temporal Logic (LTL) formulas from NL commands, grouping equivalent LTL formulas, and selecting the majority group of formulas as the final LTL formula. Constrained decoding then uses the generated LTL formula to enforce the autoregressive inference of plans, ensuring the generated plans conform to the LTL. Domain-specific fine-tuning customizes LLMs to produce safe and efficient plans within specific task domains. Our approach, Safe Efficient LLM Planner (SELP), combines these insights to create LLM planners to generate plans adhering to user commands with high confidence. We demonstrate the effectiveness and generalizability of SELP across different robot agents and tasks, including drone navigation and robot manipulation. For drone navigation tasks, SELP outperforms state-of-the-art planners by 10.8% in safety rate (i.e., finishing tasks conforming to NL commands) and by 19.8% in plan efficiency. For robot manipulation tasks, SELP achieves 20.4% improvement in safety rate. Our datasets for evaluating NL-to-LTL and robot task planning will be released in github.com/lt-asset/selp.

Manipulating Neural Path Planners via Slight Perturbations

Data-driven neural path planners are attracting increasing interest in the robotics community. However, their neural network components typically come as black boxes, obscuring their underlying decision-making processes. Their black-box nature exposes them to the risk of being compromised via the insertion of hidden malicious behaviors. For example, an attacker may hide behaviors that, when triggered, hijack a delivery robot by guiding it to a specific (albeit wrong) destination, trapping it in a predefined region, or inducing unnecessary energy expenditure by causing the robot to repeatedly circle a region. In this paper, we propose a novel approach to specify and inject a range of hidden malicious behaviors, known as backdoors, into neural path planners. Our approach provides a concise but flexible way to define these behaviors, and we show that hidden behaviors can be triggered by slight perturbations (e.g., inserting a tiny unnoticeable object), that can nonetheless significantly compromise their integrity. We also discuss potential techniques to identify these backdoors aimed at alleviating such risks. We demonstrate our approach on both sampling-based and search-based neural path planners.

Co-learning Planning and Control Policies Using Differentiable Formal Task Constraints

This paper presents a hierarchical reinforcement learning algorithm constrained by differentiable signal temporal logic. Previous work on logic-constrained reinforcement learning consider encoding these constraints with a reward function, constraining policy updates with a sample-based policy gradient. However, such techniques oftentimes tend to be inefficient because of the significant number of samples required to obtain accurate policy gradients. In this paper, instead of implicitly constraining policy search with sample-based policy gradients, we directly constrain policy search by backpropagating through formal constraints, enabling training hierarchical policies with substantially fewer training samples. The use of hierarchical policies is recognized as a crucial component of reinforcement learning with task constraints. We show that we can stably constrain policy updates, thus enabling different levels of the policy to be learned simultaneously, yielding superior performance compared with training them separately. Experiment results on several simulated high-dimensional robot dynamics and a real-world differential drive robot (TurtleBot3) demonstrate the effectiveness of our approach on five different types of task constraints. Demo videos, code, and models can be found at our project website: https://sites.google.com/view/dscrl

DistSPECTRL: Distributing Specifications in Multi-Agent Reinforcement Learning Systems

While notable progress has been made in specifying and learning objectives for general cyber-physical systems, applying these methods to distributed multi-agent systems still pose significant challenges. Among these are the need to (a) craft specification primitives that allow expression and interplay of both local and global objectives, (b) tame explosion in the state and action spaces to enable effective learning, and (c) minimize coordination frequency and the set of engaged participants for global objectives. To address these challenges, we propose a novel specification framework that allows natural composition of local and global objectives used to guide training of a multi-agent system. Our technique enables learning expressive policies that allow agents to operate in a coordination-free manner for local objectives, while using a decentralized communication protocol for enforcing global ones. Experimental results support our claim that sophisticated multi-agent distributed planning problems can be effectively realized using specification-guided learning.

Defending Observation Attacks in Deep Reinforcement Learning via Detection and Denoising

Neural network policies trained using Deep Reinforcement Learning (DRL) are well-known to be susceptible to adversarial attacks. In this paper, we consider attacks manifesting as perturbations in the observation space managed by the external environment. These attacks have been shown to downgrade policy performance significantly. We focus our attention on well-trained deterministic and stochastic neural network policies in the context of continuous control benchmarks subject to four well-studied observation space adversarial attacks. To defend against these attacks, we propose a novel defense strategy using a detect-and-denoise schema. Unlike previous adversarial training approaches that sample data in adversarial scenarios, our solution does not require sampling data in an environment under attack, thereby greatly reducing risk during training. Detailed experimental results show that our technique is comparable with state-of-the-art adversarial training approaches.

Model-free Neural Lyapunov Control for Safe Robot Navigation

Model-free Deep Reinforcement Learning (DRL) controllers have demonstrated promising results on various challenging non-linear control tasks. While a model-free DRL algorithm can solve unknown dynamics and high-dimensional problems, it lacks safety assurance. Although safety constraints can be encoded as part of a reward function, there still exists a large gap between an RL controller trained with this modified reward and a safe controller. In contrast, instead of implicitly encoding safety constraints with rewards, we explicitly co-learn a Twin Neural Lyapunov Function (TNLF) with the control policy in the DRL training loop and use the learned TNLF to build a runtime monitor. Combined with the path generated from a planner, the monitor chooses appropriate waypoints that guide the learned controller to provide collision-free control trajectories. Our approach inherits the scalability advantages from DRL while enhancing safety guarantees. Our experimental evaluation demonstrates the effectiveness of our approach compared to DRL with augmented rewards and constrained DRL methods over a range of high-dimensional safety-sensitive navigation tasks.

Scalable Synthesis of Verified Controllers in Deep Reinforcement Learning

There has been significant recent interest in devising verification techniques for learning-enabled controllers (LECs) that manage safety-critical systems. Given the opacity and lack of interpretability of the neural policies that govern the behavior of such controllers, many existing approaches enforce safety properties through the use of shields, a dynamic monitoring and repair mechanism that ensures a LEC does not emit actions that would violate desired safety conditions. These methods, however, have shown to have significant scalability limitations because verification costs grow as problem dimensionality and objective complexity increase. In this paper, we propose a new automated verification pipeline capable of synthesizing high-quality safety shields even when the problem domain involves hundreds of dimensions, or when the desired objective involves stochastic perturbations, liveness considerations, and other complex non-functional properties. Our key insight involves separating safety verification from neural controller, using pre-computed verified safety shields to constrain neural controller training which does not only focus on safety. Experimental results over a range of realistic high-dimensional deep RL benchmarks demonstrate the effectiveness of our approach.

Robustness to Adversarial Attacks in Learning-Enabled Controllers

Learning-enabled controllers used in cyber-physical systems (CPS) are known to be susceptible to adversarial attacks. Such attacks manifest as perturbations to the states generated by the controller's environment in response to its actions. We consider state perturbations that encompass a wide variety of adversarial attacks and describe an attack scheme for discovering adversarial states. To be useful, these attacks need to be natural, yielding states in which the controller can be reasonably expected to generate a meaningful response. We consider shield-based defenses as a means to improve controller robustness in the face of such perturbations. Our defense strategy allows us to treat the controller and environment as black-boxes with unknown dynamics. We provide a two-stage approach to construct this defense and show its effectiveness through a range of experiments on realistic continuous control domains such as the navigation control-loop of an F16 aircraft and the motion control system of humanoid robots.

ART: Abstraction Refinement-Guided Training for Provably Correct Neural Networks

Artificial neural networks (ANNs) have demonstrated remarkable utility in a variety of challenging machine learning applications. However, their complex architecture makes asserting any formal guarantees about their behavior difficult. Existing approaches to this problem typically consider verification as a post facto white-box process, one that reasons about the safety of an existing network through exploration of its internal structure, rather than via a methodology that ensures the network is correct-by-construction. In this paper, we present a novel learning framework that takes an important first step towards realizing such a methodology. Our technique enables the construction of provably correct networks with respect to a broad class of safety properties, a capability that goes well-beyond existing approaches. Overcoming the challenge of general safety property enforcement within the network training process in a supervised learning pipeline, however, requires a fundamental shift in how we architect and build ANNs. Our key insight is that we can integrate an optimization-based abstraction refinement loop into the learning process that iteratively splits the input space from which training data is drawn, based on the efficacy with which such a partition enables safety verification. To do so, our approach enables training to take place over an abstraction of a concrete network that operates over dynamically constructed partitions of the input space. We provide theoretical results that show that classical gradient descent methods used to optimize these networks can be seamlessly adopted to this framework to ensure soundness of our approach. Moreover, we empirically demonstrate that realizing soundness does not come at the price of accuracy, giving us a meaningful pathway for building both precise and correct networks.

An Inductive Synthesis Framework for Verifiable Reinforcement Learning

Despite the tremendous advances that have been made in the last decade on developing useful machine-learning applications, their wider adoption has been hindered by the lack of strong assurance guarantees that can be made about their behavior. In this paper, we consider how formal verification techniques developed for traditional software systems can be repurposed for verification of reinforcement learning-enabled ones, a particularly important class of machine learning systems. Rather than enforcing safety by examining and altering the structure of a complex neural network implementation, our technique uses blackbox methods to synthesizes deterministic programs, simpler, more interpretable, approximations of the network that can nonetheless guarantee desired safety properties are preserved, even when the network is deployed in unanticipated or previously unobserved environments. Our methodology frames the problem of neural network verification in terms of a counterexample and syntax-guided inductive synthesis procedure over these programs. The synthesis procedure searches for both a deterministic program and an inductive invariant over an infinite state transition system that represents a specification of an application's control logic. Additional specifications defining environment-based constraints can also be provided to further refine the search space. Synthesized programs deployed in conjunction with a neural network implementation dynamically enforce safety conditions by monitoring and preventing potentially unsafe actions proposed by neural policies. Experimental results over a wide range of cyber-physical applications demonstrate that software-inspired formal verification techniques can be used to realize trustworthy reinforcement learning systems with low overhead.