ZeroBot trains robot manipulation from scratch in 119 seconds with 87% success using generative real2sim
Ed__Johns · x · 2026-09-30
ZeroBot, a paper from Imperial College London and RAI researchers published in IEEE RA-L and presented at IROS, introduces a generative real2sim framework that learns robot manipulation from scratch in minutes — zero human demonstrations, zero policy pre-training, zero known object models.
How it works:
- Image-to-3D generative models reconstruct a full object mesh from a single view plus a goal pose
- Large-scale parallel RL runs in simulation on the generated mesh
- A novel action space leverages the generated geometry and a learned value function to sample contact-rich states, dramatically speeding up training
Evaluated on real-world grasping, pushing, articulated-object interaction and multi-stage manipulation, ZeroBot hits an 87% success rate with an average training time of just 119 seconds.
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
- Simify: training-free real-to-sim framework solves robot spatial reasoning in seconds — 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