Research Papers
Working Papers & Research
DOI-registered research outputs from the Saral Systems Council. Each paper has a permanent landing page with structured metadata, abstract, and citation information.
15 publications across energy infrastructure, robotics, AI, and quantum systems research.
Type
Domain
Showing 15 of 15 publications
Energy Theft Detection Using Graph Neural Networks on Distribution Network Topology
Sayonsom Chanda
Non-technical losses from energy theft account for 15-25% of electricity distributed in many Indian states, representing both a significant revenue gap for utilities and a barrier to grid modernization. Existing ML-based theft detection approaches treat each consumer independently, ignoring the network topology that connects them. We propose a graph neural network (GNN) approach that models the distribution network as a graph where consumers are nodes and electrical connectivity provides edges. The GNN aggregates consumption patterns from electrically adjacent consumers, enabling detection of theft patterns that are invisible when consumers are analyzed in isolation, such as coordinated meter tampering on the same lateral or theft downstream of a tampered distribution transformer. Evaluated on data from a large Indian distribution utility covering 1.2 million consumers, our approach achieves 89% precision and 76% recall for theft detection, outperforming independent consumer models by 23% in precision and 31% in recall. The paper discusses the practical considerations for deployment in Indian utility IT environments, including integration with existing billing systems and the handling of incomplete network topology data.
Bridging the Reality Gap: Domain Randomization Strategies for Indian Sewer Geometry in Isaac Sim
Sayonsom Chanda
Sim-to-real transfer for robotic navigation in standardized environments benefits from well-characterized geometry and material properties. Indian sewer infrastructure presents a distinct challenge: tunnels are hand-built with irregular cross-sections, walls are constructed from mixed materials (brick, stone, concrete, bare earth), and dimensions vary unpredictably even within a single municipal system. Standard domain randomization approaches that vary texture, lighting, and physics parameters within narrow ranges fail to capture this structural diversity. We present a geometry-aware domain randomization strategy implemented in NVIDIA Isaac Sim that procedurally generates sewer environments matching the statistical distribution of Indian underground infrastructure. Our approach randomizes tunnel cross-section shape (circular, rectangular, trapezoidal, irregular), wall material composition, surface degradation patterns, obstacle placement, and water level. We train SafAI navigation policies in these randomized environments and demonstrate a 41% improvement in zero-shot sim-to-real transfer compared to standard domain randomization when evaluated in three Indian municipal sewer systems. The procedural generation pipeline and all environment assets are released as open-source tools.
Collision-Aware Path Planning for Articulated Arms in Narrow Irregular Tunnels Under Partial Observability
Sayonsom Chanda
Path planning for articulated robotic arms in confined spaces requires reasoning about collision geometry in environments where the workspace boundary is irregular, partially observed, and may change during operation (e.g., due to water flow or debris movement). Classical motion planning algorithms assume complete workspace knowledge or rely on sensor-complete reconstruction, neither of which is practical for battery-powered robots operating in Indian sewers with limited onboard compute. We present a collision-aware planning approach that maintains a probabilistic occupancy representation of the workspace, updated incrementally from wrist-mounted depth sensors, and plans arm trajectories using a learned collision cost field. The planner operates in a receding-horizon fashion, replanning every 200ms as new depth observations arrive. Evaluated on the SafAI platform performing sludge extraction in irregular tunnel sections, our approach reduces collision events by 67% compared to a geometry-unaware baseline while maintaining 91% of the baseline's task completion rate. The method requires less than 15% of available onboard compute, leaving sufficient headroom for simultaneous VLA model inference.
Embodiment-Aware Intent Transfer for Hazardous Labour Substitution: From Simulation to Indian Field Conditions
Sayonsom Chanda
Transfer learning for robotic manipulation typically assumes a consistent sim-to-real gap characterized by visual and dynamics differences between simulator and target environment. We identify a distinct and underexplored transfer challenge in hazardous labour substitution: embodiment-aware intent transfer, where the source demonstrations come from human workers performing tasks with their bodies, and the target is a robotic system with fundamentally different morphology. In the context of Indian sewer maintenance, workers use their hands, feet, and improvised tools to navigate, extract, and deposit material in irregular confined spaces. We propose a framework that decomposes human task demonstrations into embodiment-invariant intent representations, which are then re-grounded in the robot's action space through a learned mapping. We evaluate this approach using data from our ethical wearable data collection programme with Indian sanitation workers and demonstrate a 34% improvement in task completion over direct behavioral cloning from human demonstrations. The paper discusses the broader implications of intent transfer for deploying robots in labor contexts where the incumbent workers, not engineers, are the domain experts.
India Data Center Review 2026: Regional Energy Demand Outlook
Multiple Authors
The India Data Center Review 2026 presents the first comprehensive, independently produced analysis of energy demand projections for India's rapidly expanding data center sector. Drawing on primary infrastructure surveys, utility filings, and satellite imagery analysis, this review maps current and planned data center capacity across seven major clusters: Mumbai, Chennai, Hyderabad, Pune, Delhi-NCR, Bengaluru, and Kolkata. We project total data center power demand to reach 4.5-5.2 GW by 2028, representing a compound annual growth rate of 28-33% from the 2024 baseline. The review examines the grid integration challenges posed by this growth, including transmission constraints, renewable energy procurement pathways, and the emerging tension between hyperscaler sustainability commitments and the carbon intensity of India's regional grids. Each regional chapter includes a power demand forecast, grid capacity assessment, and policy landscape analysis. The review is published under the India Data Center Review ISSN and is intended as an annual reference for infrastructure planners, investors, and policymakers.
Adaptive Precision Inference for Battery-Powered Field Robots in Infrastructure-Poor Environments
Sayonsom Chanda
Battery-powered robots deployed in infrastructure-poor environments face a fundamental tradeoff between inference quality and operational endurance. Static quantization schemes, which fix precision at deployment time, cannot adapt to the varying computational demands of different task phases. We present an adaptive precision inference framework that dynamically adjusts model quantization level based on real-time battery state, task criticality, and environmental complexity. The framework monitors three signals: remaining battery charge, a learned task-phase classifier that distinguishes high-criticality phases (e.g., extraction near a pipe junction) from low-criticality phases (e.g., straight-line navigation), and a visual complexity estimator. During low-criticality phases, the system aggressively quantizes to INT4, while escalating to FP16 during critical manipulation sequences. Evaluated on the SafAI sewer inspection platform, adaptive precision extends operational time by 2.1x compared to fixed FP16 inference while maintaining 96% of full-precision task success rate. We release the framework as an open-source library compatible with ONNX Runtime and TensorRT.
Quantum-in-the-Loop Architecture for Real-Time Grid Contingency Analysis
Sayonsom Chanda
Real-time contingency analysis in power systems requires evaluating thousands of potential failure scenarios within the few minutes available for operator decision-making. As grid complexity increases with distributed energy resources and bidirectional power flows, the computational burden of N-1 and N-2 contingency screening grows beyond the capability of current classical screening methods. We propose a quantum-in-the-loop architecture that integrates quantum computing resources into the real-time operations environment of a grid control center. The architecture uses classical pre-screening to identify the most computationally demanding contingencies, routes these to a quantum processor for rapid evaluation using a variational quantum eigensolver adapted for power flow equations, and returns results to the classical energy management system within the operational time window. We evaluate the architecture in simulation using a digital twin of a 500-bus regional transmission system and demonstrate that quantum-assisted contingency screening can evaluate 3.1x more N-2 contingencies within a 5-minute operational window compared to a purely classical approach. The paper discusses the integration challenges with existing EMS/SCADA systems, the latency requirements for quantum cloud access, and a phased deployment roadmap for Indian transmission operators.
SmolVLA-SewerBot: A Compact Vision-Language-Action Model for Autonomous Sewer Inspection in Resource-Constrained Environments
Sayonsom Chanda
We introduce SmolVLA-SewerBot, a compact vision-language-action model optimized for deployment on battery-powered robotic platforms operating in confined underground spaces. The model architecture extends the SmolVLA framework with task-specific adaptations for sewer inspection: a dual-camera visual encoder processing simultaneous front and wrist camera feeds, a language-conditioned action decoder supporting natural language task specifications in Hindi and English, and an adaptive precision inference engine that dynamically adjusts quantization levels based on available battery and computational headroom. Trained on 12,000 episodes of teleoperated sewer inspection data collected from Indian municipal systems, the model achieves 87% task completion rate across four canonical sewer maintenance subtasks (navigate, assess, extract, deposit) when deployed on a Jetson Orin Nano. We benchmark against full-precision baselines and show that our adaptive quantization approach maintains 94% of baseline accuracy while reducing inference energy consumption by 3.8x, extending operational time from 45 minutes to 2.7 hours on a single battery charge. The model weights and evaluation code are released on HuggingFace.
Edge AI for Appliance Identification and Energy Disaggregation on Indian Load Profiles
Sayonsom Chanda
Non-intrusive load monitoring (NILM) research has predominantly focused on load profiles from North American and European households, where appliance mixes, usage patterns, and electrical characteristics differ substantially from Indian consumption. This paper presents edge-deployable NILM models trained and evaluated on Indian residential and small-commercial load profiles. We collected a 14-month dataset from 230 premises across four Indian cities, capturing appliance-level ground truth through a combination of smart plugs and manual annotation. Our lightweight convolutional architecture, designed for deployment on ARM Cortex-M microcontrollers in smart meters, achieves 82% F1 score for appliance identification across 12 common Indian appliance categories, including several (desert coolers, water pumps, inverter-battery systems) absent from standard NILM benchmarks. We show that models trained on Western datasets achieve only 54% F1 when applied to Indian load profiles without fine-tuning, confirming the need for India-specific training data. The model, training code, and a subset of the anonymized dataset are released as open-source resources.
Vision-Language-Action Models for Sewer Inspection Robotics in India
Sayonsom Chanda
We present SafAI, India's first indigenously trained vision-language-action (VLA) model designed for autonomous sewer inspection in non-standardized underground infrastructure. Manual scavenging, despite being legally prohibited since 2013, continues to claim over 400 lives annually in India. Existing robotic solutions, designed for standardized Western sewer systems, fail within meters of deployment in Indian conditions characterized by irregular brick-lined tunnels, open drains, and unmapped septic tanks. SafAI addresses this gap through a compact VLA architecture (SmolVLA-SewerBot) trained on field data collected from Indian sanitation workers wearing sensor arrays under an ethical consent framework. The model achieves autonomous navigation, blockage assessment, sludge extraction, and material deposition on battery-powered hardware costing under INR 2 lakh. We report results from simulation in Isaac Sim with domain randomization calibrated to Indian sewer geometry, and from initial field trials in municipal drain systems. The paper also describes our data collection methodology, which compensates participating workers and ensures informed consent at every stage.
Ethical Wearable Data Collection from Sanitation Workers: A Consent Framework for Robotics Training Corpora
Sayonsom Chanda
Training robotic systems to perform hazardous tasks currently done by human workers requires demonstration data from those workers. This creates an ethical tension: the workers whose labor generates the training data are the same population whose livelihoods may be affected by the resulting automation. We present a consent and compensation framework developed through two years of fieldwork with sanitation workers in Indian municipalities. The framework addresses four dimensions: informed consent in workers' native languages with visual documentation of data usage; fair compensation that exceeds daily wages and includes participation in downstream value; data sovereignty provisions that give workers and their cooperatives control over how their data is used; and transition support connecting workers to alternative employment. We report on the framework's implementation across three field campaigns involving 47 workers, document the challenges of operationalizing ethical data collection in informal labor contexts, and propose standards for the emerging field of human-to-robot knowledge transfer. The framework is published as an open protocol for other robotics research groups working in labor substitution contexts.
KTBot: Preserving Tribal Knowledge in Industrial Operations Through Conversational AI
Sayonsom Chanda
Critical industrial facilities rely on tacit knowledge accumulated by experienced operators over decades of hands-on work. This knowledge, often called tribal knowledge, encompasses diagnostic heuristics, exception-handling procedures, and contextual judgment that is rarely documented in standard operating procedures. As experienced operators retire, this knowledge is permanently lost, degrading operational reliability and safety. We present KTBot, a conversational AI system designed to capture, structure, and make queryable the tribal knowledge of industrial operators. KTBot uses a structured interview protocol to elicit knowledge through scenario-based conversations, stores the resulting knowledge in a graph-structured representation that preserves causal relationships and contextual conditions, and makes it accessible through a natural language query interface. We report on a deployment at a thermal power plant where KTBot captured knowledge from 12 operators with a combined 340 years of experience, covering 847 distinct operational scenarios. New engineers using KTBot resolved unfamiliar operational situations 2.4x faster than those relying on documentation alone, with a 67% reduction in escalations to senior operators. The paper discusses the design principles for knowledge elicitation from domain experts who are not technology-fluent, and the organizational dynamics of deploying knowledge preservation systems in unionized industrial environments.
Hybrid Quantum-Classical Optimization for 118-Bus Power Systems
Sayonsom Chanda
This paper demonstrates a hybrid quantum-classical optimization approach for solving the optimal power flow (OPF) problem on the IEEE 118-bus test system. Classical OPF solvers scale poorly with network size and the combinatorial complexity introduced by discrete control variables such as transformer tap positions and switched shunt capacitors. We formulate a decomposed optimization where a quantum approximate optimization algorithm (QAOA) handles the discrete subproblem while a classical interior-point method solves the continuous relaxation. Using IBM's Qiskit runtime on a 127-qubit Eagle processor, we benchmark solution quality and computation time against purely classical solvers (IPOPT, Gurobi) and find that the hybrid approach achieves comparable solution quality with a 2.3x reduction in computation time for the discrete subproblem on systems with more than 50 discrete variables. We discuss the practical implications for real-time grid operations, the current limitations of quantum hardware noise, and a roadmap for scaling to transmission networks with thousands of buses. The paper includes reproducible code and benchmark datasets published through the Saral Systems Council data archive.
Robotic Systems and Manual Scavenging in India: A Five-Year Research Agenda for Zero Human Entry
Sayonsom Chanda
Despite legal prohibition, manual scavenging persists across India due to a combination of regulatory gaps, economic constraints, and the absence of affordable robotic alternatives capable of operating in India's non-standardized underground infrastructure. This paper presents a five-year research agenda for achieving zero human entry in Indian sewer and septic tank maintenance. The agenda is organized around four research pillars: (1) developing affordable, edge-deployed VLA models trained on Indian sewer geometry; (2) building robust sim-to-real transfer pipelines that account for the extreme variability of hand-built underground infrastructure; (3) establishing ethical data collection practices that compensate sanitation workers as domain experts; and (4) creating policy frameworks that align municipal procurement with robotic capability timelines. We draw on SARAL's ongoing SafAI programme to ground each pillar in current technical capabilities and identify specific research gaps. The paper argues that the primary barrier to robotic sewer maintenance in India is not fundamental AI capability but rather the absence of India-specific training data, the cost structure of existing robotic platforms, and institutional inertia in municipal procurement. We propose specific milestones, resource requirements, and partnership structures needed to close these gaps within five years.
India's Submarine Cable Infrastructure and Data Center Geography
Sayonsom Chanda
This paper maps the relationship between submarine cable landing stations and the emerging geography of data center clusters across India. Drawing on primary geospatial data collected by the Saral Systems Council, we identify thirteen landing stations across six coastal states and analyze their proximity to planned and operational data center campuses. The analysis reveals that India's data center geography is increasingly shaped by international bandwidth availability rather than domestic demand patterns alone. We find that landing stations in Mumbai and Chennai anchor over 78% of India's international bandwidth capacity, creating a coastal concentration that has significant implications for latency, resilience, and regional digital equity. The paper proposes a connectivity-weighted site selection framework for data center developers and policymakers seeking to balance infrastructure efficiency with geographic distribution.
Citation
How to Cite
Each paper landing page includes a pre-formatted citation. Use the following general format for Saral Systems Council working papers:
Author (Year). "Paper Title." Saral Systems Council Working Paper SSC-WP-YYYY-NNN. DOI: 10.xxxx/ssc-wp-yyyy-nnn