GenOSIL: Generalized Optimal and Safe Robot Control using Parameter-Conditioned Imitation Learning

Mumuksh Tayal, Manan Tayal, Ravi Prakrash

Cyber Physical Systems, Indian Institute of Science, Bangalore

GenOSIL Framework

GenOSIL Architecture: (Left) During training, expert demonstration trajectories—augmented with obstacle parameters, and goal coordinates—are used to learn the Variational Encoder to produce a latent representation (z), which is combined with agent’s state to drive training of the Policy Network. (Right) At inference, trained architecture deploys the Policy Network on a real agent, generating actions for safe navigation in dynamic environments.

Autonomous Ground Vehicle

Franka Manipulator task

Hardware Demonstrations

Demonstration 1

Demonstration 2

Comparison

BibTeX


      @misc{tayal2025genosilgeneralizedoptimalsafe,
        title={GenOSIL: Generalized Optimal and Safe Robot Control using Parameter-Conditioned Imitation Learning}, 
        author={Mumuksh Tayal and Manan Tayal and Ravi Prakash},
        year={2025},
        eprint={2503.12243},
        archivePrefix={arXiv},
        primaryClass={cs.RO},
        url={https://arxiv.org/abs/2503.12243}, 
  }