Research I've written, reviewed, and published.

Uncertainty-Aware Domain Randomisation for Robust Sim-to-Real Transfer in Dexterous Grasping

E. ARI, K. Lindqvist, S. Öztürk

We present an uncertainty-aware domain randomisation framework that adaptively schedules simulation parameters during policy training based on estimated uncertainty in the physics model. Our approach reduces the sim-to-real gap by 34% compared to fixed-schedule randomisation, as measured by grasp success rate on a UR5e platform across 40 unseen object geometries. We further show that the trained policies exhibit interpretable uncertainty representations that correlate with out-of-distribution contact scenarios.

Kepler: A Modular Benchmark for Motion Planning in Constrained Environments

E. ARI, B. Chen, F. Kılıçlı

We introduce Kepler, an open-source benchmarking suite for motion planning algorithms in robotics. Kepler provides 12 standardised task environments ranging from free-space manipulation to highly constrained industrial assembly scenarios, unified metrics for comparing planners, and baseline results for six widely-used algorithms. We demonstrate significant performance disparities across environments that are obscured by aggregate metrics used in prior comparisons.

Sample-Efficient Sim-to-Real Transfer for Dexterous Manipulation via Hierarchical Policy Decomposition

E. ARI, M. Schneider, H. van der Berg, A. Yılmaz

We propose a hierarchical policy architecture for dexterous manipulation that decomposes the task into a high-level grasp planner and a low-level contact controller. This decomposition dramatically reduces the sample complexity of sim-to-real transfer by localising the physics gap to the contact controller, which can be fine-tuned with as few as 200 real-world demonstrations. Evaluated on a UR5e arm with a two-finger gripper across 45 object categories from the YCB dataset.

Adaptive Gait Synthesis for Quadruped Locomotion on Heterogeneous Terrain using Learned Terrain Embeddings

M. Hoffmann, E. ARI, R. Bhattacharya, L. Köhler

We present a terrain-aware quadruped locomotion framework that uses a convolutional encoder to extract compact terrain embeddings from height maps, which are then used to condition a locomotion policy. The framework enables rapid adaptation to novel terrain types without retraining the base policy. Validated on ANYmal-C across 8 outdoor terrain categories including gravel, mud, steps, and ramps.

A Survey of Sim-to-Real Transfer Methods for Robotic Manipulation: Challenges, Benchmarks, and Open Problems

E. ARI, S. Park, C. Martinez, D. Fischer

We present a comprehensive survey of sim-to-real transfer methods in robotic manipulation, covering domain randomisation, system identification, adaptive control, and hybrid approaches. We identify three fundamental challenges — physics accuracy, sensor simulation, and contact modelling — and propose a taxonomy for classifying existing methods along these dimensions. We conclude with a structured benchmark proposal and a set of open problems that we believe are underexplored in the current literature.

Deep Defect Detection at the Edge: Efficient EfficientDet Variants for Industrial Visual Inspection

E. ARI, T. Müller, S. Özkan

We systematically evaluate a family of EfficientDet variants for industrial defect detection under strict latency and memory constraints on NVIDIA Jetson AGX hardware. We identify model configurations that achieve state-of-the-art accuracy on three industrial inspection datasets while maintaining 40ms inference latency, and present a practical guide for deploying such systems in manufacturing environments.

Residual Physics Learning for Contact-Rich Manipulation Tasks

A. Kim, E. ARI, N. Dubowsky

Contact-rich manipulation tasks challenge physics simulators because of unmodelled contact dynamics. We propose learning a residual physics model that captures the discrepancy between simulator and real-world contact behaviour. Combined with a model-based planner, our approach achieves reliable peg-in-hole insertion across tolerances as small as 0.3mm without any task-specific fine-tuning on the real robot.

Online Adaptation of Motion Planning Heuristics Using Execution Experience

E. ARI, P. Weston

Motion planners are typically configured offline with fixed heuristics that do not adapt to the specific environment or robot configuration encountered at runtime. We propose a lightweight online adaptation mechanism that updates planner heuristics based on recent execution experience, improving planning time by 62% in repeated navigation tasks without sacrificing path quality.

Curriculum-Based Domain Randomisation for Legged Locomotion in Structured Environments

E. ARI, R. Grandia, M. Hutter

Force-Torque Sensor Fusion for Reactive Grasping in Unstructured Environments

K. Lindqvist, E. ARI, G. Allibert

Domain Randomisation Strategies for Legged Locomotion: An Empirical Analysis

E. ARI, M. Hutter, R. Grandia, F. Jenelten

Periodic Gait Controllers for Quadruped Locomotion via Constrained Reinforcement Learning

E. ARI, F. Jenelten, M. Hutter

Adaptive Locomotion over Rough Terrain Using a Learned Elevation Map Representation

R. Grandia, E. ARI, M. Hutter

Model Predictive Control for Quadrotor UAV Navigation with Dynamically Reconfigurable Constraints

E. ARI, K. Altun