Chayan Banerjee

Chayan Banerjee, Kien Nguyen, Clinton Fookes et al.

The inclusion of physical information in machine learning frameworks has revolutionized many application areas. This involves enhancing the learning process by incorporating physical constraints and adhering to physical laws. In this work we explore their utility for reinforcement learning applications. We present a thorough review of the literature on incorporating physics information, as known as physics priors, in reinforcement learning approaches, commonly referred to as physics-informed reinforcement learning (PIRL). We introduce a novel taxonomy with the reinforcement learning pipeline as the backbone to classify existing works, compare and contrast them, and derive crucial insights. Existing works are analyzed with regard to the representation/ form of the governing physics modeled for integration, their specific contribution to the typical reinforcement learning architecture, and their connection to the underlying reinforcement learning pipeline stages. We also identify core learning architectures and physics incorporation biases (i.e., observational, inductive and learning) of existing PIRL approaches and use them to further categorize the works for better understanding and adaptation. By providing a comprehensive perspective on the implementation of the physics-informed capability, the taxonomy presents a cohesive approach to PIRL. It identifies the areas where this approach has been applied, as well as the gaps and opportunities that exist. Additionally, the taxonomy sheds light on unresolved issues and challenges, which can guide future research. This nascent field holds great potential for enhancing reinforcement learning algorithms by increasing their physical plausibility, precision, data efficiency, and applicability in real-world scenarios.

Chayan Banerjee, Ethan Goan

Deploying reinforcement learning in safety critical domains, from autonomous vehicles to medical decision support, is constrained by failures arising when systems encounter unfamiliar conditions. We argue that the fundamental bottleneck is not individual challenges like changing dynamics or incomplete observations, but their synergistic interaction, which we term the Epistemic Trap: agents cannot estimate their state without knowing system dynamics, nor learn dynamics without accurate state information. Proof-of-concept experiments in simulated locomotion reveal that combining these uncertainties causes failures far worse than either challenge alone, a 77% performance degradation against the 46% by adding the individual effects, demonstrating compounding failure modes that conventional methods overlook. Such approaches adopt a passive epistemic stance that cannot resolve this coupled uncertainty. We propose reframing safety as an information problem, introducing an Adaptive Safety Architecture built around three contributions: the Compound Uncertainty Coefficient ($κ$), a mutual information based metric that quantifies state dynamics coupling and is computable online without full joint belief inference; information seeking policies governed by a MaxInfoRL objective that actively probe system dynamics; and regime-adaptive safety constraints that tighten as epistemic coupling rises. This paradigm shift, from passive robustness to active perception, offers a principled path toward decision making systems that operate under uncertainty, recognize their own ignorance, and act strategically to resolve it.

Chayan Banerjee, Zhiyong Chen, Nasimul Noman

Actor-critic (AC) algorithms are a class of model-free deep reinforcement learning algorithms, which have proven their efficacy in diverse domains, especially in solving continuous control problems. Improvement of exploration (action entropy) and exploitation (expected return) using more efficient samples is a critical issue in AC algorithms. A basic strategy of a learning algorithm is to facilitate indiscriminately exploring all of the environment state space, as well as to encourage exploring rarely visited states rather than frequently visited one. Under this strategy, we propose a new method to boost exploration through an intrinsic reward, based on measurement of a state's novelty and the associated benefit of exploring the state (with regards to policy optimization), altogether called plausible novelty. With incentivized exploration of plausible novel states, an AC algorithm is able to improve its sample efficiency and hence training performance. The new method is verified by extensive simulations of continuous control tasks of MuJoCo environments on a variety of prominent off-policy AC algorithms.

Chayan Banerjee, Kien Nguyen, Olivier Salvado et al.

The incorporation of physical information in machine learning frameworks is transforming medical image analysis (MIA). By integrating fundamental knowledge and governing physical laws, these models achieve enhanced robustness and interpretability. In this work, we explore the utility of physics-informed approaches for MIA (PIMIA) tasks such as registration, generation, classification, and reconstruction. We present a systematic literature review of over 80 papers on physics-informed methods dedicated to MIA. We propose a unified taxonomy to investigate what physics knowledge and processes are modelled, how they are represented, and the strategies to incorporate them into MIA models. We delve deep into a wide range of image analysis tasks, from imaging, generation, prediction, inverse imaging (super-resolution and reconstruction), registration, and image analysis (segmentation and classification). For each task, we thoroughly examine and present in a tabular format the central physics-guided operation, the region of interest (with respect to human anatomy), the corresponding imaging modality, the dataset used for model training, the deep network architecture employed, and the primary physical process, equation, or principle utilized. Additionally, we also introduce a novel metric to compare the performance of PIMIA methods across different tasks and datasets. Based on this review, we summarize and distil our perspectives on the challenges, open research questions, and directions for future research. We highlight key open challenges in PIMIA, including selecting suitable physics priors and establishing a standardized benchmarking platform.

Chayan Banerjee, Kien Nguyen, Clinton Fookes

Optimizing the mining process -- particularly truck dispatch scheduling -- is a key driver of efficiency in open-pit operations. However, the dynamic and stochastic nature of these environments, with uncertainties such as equipment failures, truck maintenance, and variable haul cycle times, challenges traditional optimization. While Reinforcement Learning (RL) shows strong potential for adaptive decision-making in mining logistics, practical deployment requires evaluation in realistic, customizable simulation environments. The lack of standardized benchmarking hampers fair algorithm comparison, reproducibility, and real-world applicability of RL solutions. To address this, we present Mining-Gym -- a configurable, open-source benchmarking environment for training, testing, and evaluating RL algorithms in mining process optimization. Built on Salabim-based Discrete Event Simulation (DES) and integrated with Gymnasium, Mining-Gym captures mining-specific uncertainties through an event-driven decision-point architecture. It offers a GUI for parameter configuration, data logging, and real-time visualization, supporting reproducible evaluation of RL strategies and heuristic baselines. We validate Mining-Gym by comparing classical heuristics with RL-based scheduling across six scenarios from normal operation to severe equipment failures. Results show it is an effective, reproducible testbed, enabling fair evaluation of adaptive decision-making and demonstrating the strong performance potential of RL-trained schedulers.

Aayush Agrawal, Ashmitha Jaysi Sivakumar, Ibrahim Kaif et al.

Accurate lane detection is essential for effective path planning and lane following in autonomous driving, especially in scenarios with significant occlusion from vehicles and pedestrians. Existing models often struggle under such conditions, leading to unreliable navigation and safety risks. We propose two innovative approaches to enhance lane detection in these challenging environments, each showing notable improvements over current methods. The first approach aug-Segment improves conventional lane detection models by augmenting the training dataset of CULanes with simulated occlusions and training a segmentation model. This method achieves a 12% improvement over a number of SOTA models on the CULanes dataset, demonstrating that enriched training data can better handle occlusions, however, since this model lacked robustness to certain settings, our main contribution is the second approach, LOID Lane Occlusion Inpainting and Detection. LOID introduces an advanced lane detection network that uses an image processing pipeline to identify and mask occlusions. It then employs inpainting models to reconstruct the road environment in the occluded areas. The enhanced image is processed by a lane detection algorithm, resulting in a 20% & 24% improvement over several SOTA models on the BDDK100 and CULanes datasets respectively, highlighting the effectiveness of this novel technique.

Chayan Banerjee

A reliable action feasibility scorer is a critical bottleneck in embodied agent pipelines: before any planning or reasoning occurs, the agent must identify which candidate actions are physically executable in the current state. Existing approaches use supervised fine-tuning (SFT) to train action scorers, but SFT treats each candidate independently and does not explicitly teach the model to discriminate between actions that are physically correct and those that are subtly wrong. We propose the Contrastive World Model (CWM), which fine-tunes a large language model (LLM) as an action scorer using an InfoNCE contrastive objective with hard-mined negative examples. The key idea is to push valid actions away from invalid ones in scoring space, with special emphasis on hard negatives: semantically similar but physically incompatible candidates. We evaluate CWM on the ScienceWorld benchmark through two studies. First, an intrinsic affordance evaluation on 605 hard-negative test pairs shows that CWM outperforms SFT by +6.76 percentage points on Precision@1 for minimal-edit negatives -- cases where a single word changes the physical outcome -- and achieves a higher AUC-ROC (0.929 vs. 0.906). Second, a live filter characterisation study measures how well CWM ranks gold-path actions against all valid environment actions during task execution. Under out-of-distribution stress conditions, CWM maintains a significantly better safety margin (-2.39) than SFT (-3.96), indicating that the gold action is ranked closer to the top. These results support the hypothesis that contrastive training induces representations that capture physical feasibility more faithfully than SFT alone.

Mohammad Saif Nazir, Chayan Banerjee

Reinforcement learning (RL) often struggles with reward misalignment, where agents optimize given rewards but fail to exhibit the desired behaviors. This arises when the reward function incentivizes proxy behaviors misaligned with the true objective. While human-in-the-loop (HITL) methods can mitigate this issue, they also introduce biases, leading to inconsistent and subjective feedback that complicates learning. To address these challenges, we propose two key contributions. First, we extend the use of zero-shot, off-the-shelf large language models (LLMs) for reward shaping beyond natural language processing (NLP) to continuous control tasks. Using LLMs as direct feedback providers eliminates the need for surrogate models trained on human feedback, which often inherit biases from training data. Second, we introduce a hybrid framework (LLM-HFBF) that enables LLMs to identify and correct biases in human feedback while incorporating this feedback into the reward shaping process. The LLM-HFBF framework creates a more balanced and reliable system by addressing both the limitations of LLMs (e.g., lack of domain-specific knowledge) and human supervision (e.g., inherent biases). By enabling human feedback bias flagging and correction, our approach improves reinforcement learning performance and reduces reliance on potentially biased human feedback. Empirical experiments show that biased human feedback significantly reduces performance, with Average Episodic Reward dropping by nearly 94% compared to unbiased approaches. In contrast, LLM-based methods sustain performance at a similar level to unbiased feedback, even in challenging edge-case scenarios.

Anirud Nandakumar, Chayan Banerjee, Lelitha Devi Vanajakshi

Efficient traffic signal control (TSC) is crucial for reducing congestion, travel delays, pollution, and for ensuring road safety. Traditional approaches, such as fixed signal control and actuated control, often struggle to handle dynamic traffic patterns. In this study, we propose a novel adaptive TSC framework that leverages Reinforcement Learning (RL), using the Proximal Policy Optimization (PPO) algorithm, to minimize total queue lengths across all signal phases. The challenge of efficiently representing highly stochastic traffic conditions for an RL controller is addressed through multiple state representations, including an expanded state space, an autoencoder representation, and a K-Planes-inspired representation. The proposed algorithm has been implemented using the Simulation of Urban Mobility (SUMO) traffic simulator and demonstrates superior performance over both traditional methods and other conventional RL-based approaches in reducing queue lengths. The best performing configuration achieves an approximately 29% reduction in average queue lengths compared to the traditional Webster method. Furthermore, comparative evaluation of alternative reward formulations demonstrates the effectiveness of the proposed queue-based approach, showcasing the potential for scalable and adaptive urban traffic management.

Ryan Chappell, Chayan Banerjee, Kien Nguyen et al.

Accurate modeling of personalized cardiovascular dynamics is crucial for non-invasive monitoring and therapy planning. State-of-the-art physics-informed neural network (PINN) approaches employ deep, multi-branch architectures with adversarial or contrastive objectives to enforce partial differential equation constraints. While effective, these enhancements introduce significant training and implementation complexity, limiting scalability and practical deployment. We investigate physics-informed neural operator learning models as efficient supervisory signals for training simplified architectures through knowledge distillation. Our approach pre-trains a physics-informed DeepONet (PI-DeepONet) on high-fidelity cuffless blood pressure recordings to learn operator mappings from raw wearable waveforms to beat-to-beat pressure signals under embedded physics constraints. This pre-trained operator serves as a frozen supervisor in a lightweight knowledge-distillation pipeline, guiding streamlined base models that eliminate complex adversarial and contrastive learning components while maintaining performance. We characterize the role of physics-informed regularization in operator learning and demonstrate its effectiveness for supervisory guidance. Through extensive experiments, our operator-supervised approach achieves performance parity with complex baselines (correlation: 0.766 vs. 0.770, RMSE: 4.452 vs. 4.501), while dramatically reducing architectural complexity from eight critical hyperparameters to a single regularization coefficient and decreasing training overhead by 4%. Our results demonstrate that operator-based supervision effectively replaces intricate multi-component training strategies, offering a more scalable and interpretable approach to physiological modeling with reduced implementation burden.

Chayan Banerjee, Kien Nguyen, Clinton Fookes et al.

The incorporation of physical information in machine learning frameworks is opening and transforming many application domains. Here the learning process is augmented through the induction of fundamental knowledge and governing physical laws. In this work, we explore their utility for computer vision tasks in interpreting and understanding visual data. We present a systematic literature review of more than 250 papers on formulation and approaches to computer vision tasks guided by physical laws. We begin by decomposing the popular computer vision pipeline into a taxonomy of stages and investigate approaches to incorporate governing physical equations in each stage. Existing approaches in computer vision tasks are analyzed with regard to what governing physical processes are modeled and formulated, and how they are incorporated, i.e. modification of input data (observation bias), modification of network architectures (inductive bias), and modification of training losses (learning bias). The taxonomy offers a unified view of the application of the physics-informed capability, highlighting where physics-informed learning has been conducted and where the gaps and opportunities are. Finally, we highlight open problems and challenges to inform future research. While still in its early days, the study of physics-informed computer vision has the promise to develop better computer vision models that can improve physical plausibility, accuracy, data efficiency, and generalization in increasingly realistic applications.

Chayan Banerjee, Zhiyong Chen, Nasimul Noman

Soft Actor-Critic (SAC) is an off-policy actor-critic reinforcement learning algorithm, essentially based on entropy regularization. SAC trains a policy by maximizing the trade-off between expected return and entropy (randomness in the policy). It has achieved state-of-the-art performance on a range of continuous-control benchmark tasks, outperforming prior on-policy and off-policy methods. SAC works in an off-policy fashion where data are sampled uniformly from past experiences (stored in a buffer) using which parameters of the policy and value function networks are updated. We propose certain crucial modifications for boosting the performance of SAC and make it more sample efficient. In our proposed improved SAC, we firstly introduce a new prioritization scheme for selecting better samples from the experience replay buffer. Secondly we use a mixture of the prioritized off-policy data with the latest on-policy data for training the policy and the value function networks. We compare our approach with the vanilla SAC and some recent variants of SAC and show that our approach outperforms the said algorithmic benchmarks. It is comparatively more stable and sample efficient when tested on a number of continuous control tasks in MuJoCo environments.

Chayan Banerjee, Zhiyong Chen, Nasimul Noman et al.

Actor-critic (AC) algorithms are known for their efficacy and high performance in solving reinforcement learning problems, but they also suffer from low sampling efficiency. An AC based policy optimization process is iterative and needs to frequently access the agent-environment system to evaluate and update the policy by rolling out the policy, collecting rewards and states (i.e. samples), and learning from them. It ultimately requires a huge number of samples to learn an optimal policy. To improve sampling efficiency, we propose a strategy to optimize the training dataset that contains significantly less samples collected from the AC process. The dataset optimization is made of a best episode only operation, a policy parameter-fitness model, and a genetic algorithm module. The optimal policy network trained by the optimized training dataset exhibits superior performance compared to many contemporary AC algorithms in controlling autonomous dynamical systems. Evaluation on standard benchmarks show that the method improves sampling efficiency, ensures faster convergence to optima, and is more data-efficient than its counterparts.