Sim-to-Real-Transfer-Risk-Dataset-Robotics-RL

Simulated in-sim training curves and real-world deployment outcomes for 750 reinforcement learning sim-to-real configurations across five India-relevant robotics tasks, grounded in domain randomization, rapid motor adaptation, massively parallel simulation, and 2024 LLM-guided sim-to-real transfer research. Forward-looking task predicts real-world deployment success from sim-only, half-trained checkpoints.

About Dataset

This dataset simulates the relationship between simulation-training choices and real-world deployment outcomes for RL policies, across 750 configurations spanning five India-relevant robotics tasks (agricultural drone spraying, warehouse navigation, contact-rich manipulation, quadruped locomotion, low-speed autonomous e-rickshaw maneuvering), five randomization methods, parallel-environment counts from 64 to 16384, and offline pretraining choices. Every row is a simulation-phase observable; the real-world deployment label is never directly observable during simulation, by construction. Ten grounding sources span 2017 to 2024, documented in sim2real_theory.py and the README Research basis.

Purpose of Dataset

Lets Compute-constrained Robotics Teams Estimate, From Cheap Simulation-only Signals, Whether A Given Sim-to-real Configuration Is Likely To Transfer Successfully To Real Hardware Before Committing Scarce Real-world Trial Budget.