Simify: training-free real-to-sim framework solves robot spatial reasoning in seconds
Ed__Johns · x · 2026-09-30
Ivan Kapelyukh's PhD capstone paper Simify (IROS 2026), from Edward Johns' lab, introduces a training-free, test-time framework for robot spatial reasoning:
- Real2Sim from one image: reconstructs simulation-ready assets from a single RGB-D photo using 3D generative models and vision-language models.
- Massively parallel evolutionary search: given a reward function (e.g., build the tallest tower), it launches thousands of parallel rollouts and evolves object arrangements, typically converging within seconds.
- Real-robot validation: quantitative experiments show end-to-end object rearrangement with unseen objects, outperforming prior foundation-model approaches for spatial reasoning.
- The paper highlights that complete, accurate geometry is critical for sim-to-real transfer.
More from Embodied
- AnyStep-WAM Cuts Denoising Steps by Up to 85% in World Action Models Without Losing Success Rate — Rui Wang · 2026-09-30
- ZeroBot trains robot manipulation from scratch in 119 seconds with 87% success using generative real2sim — Ed__Johns · 2026-09-30
- LadderMan, zero-shot sim-to-real humanoid ladder climbing, wins CoRL 2026 Spotlight — yuewang314 · 2026-09-30
- Reddit users warm to Goose, a non-humanoid home robot that fetches and grabs — fhpapa · 2026-09-30
- Amazon's StructRL uses verifiable intermediate rewards to improve long-horizon robot manipulation — amazon · 2026-09-30
- Memory price surge threatens on-device AI hardware: BOM tops $300, device may cost $1,500 — stephen280up · 2026-09-30