Reproduce “iARCS: Iterative Agentic RL for Controllable 3D Scene Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Saugat Adhikari, Ashok Prasad Neupane, Pramish Paudel, Ajad Chhatkuli, Danda Pani Paudel
iARCS adapts a pretrained indoor-scene generator to produce layouts that meet task-specific geometric requirements. An LLM converts prompts into measurable constraints and executable rewards, then reinforcement learning fine-tunes a separate LoRA adapter for each constraint. For robot-grasping, 73.83% of generated scenes pass the reachability check, versus 4.53% for MiDiffusion; this is a geometric layout test, not physical robot performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.