CoRL 2026: Instant-Fold Learns Deformable Object Manipulation from a Single Human Demo
Ed__Johns · x · 2026-11-11
Yilong Wang, Cheng Qian, and Edward Johns (Imperial College) will present Instant-Fold at CoRL 2026 in Austin on Nov 10. The in-context imitation learning framework for deformable object manipulation:
- Given a single human demonstration as prompt, it infers and executes diverse manipulation modes (spatial/ordering variations) without gradient updates.
- Method: temporal contrastive pretraining lifts masked RGB-D observations into geo-semantic 3D cloth tokens; a flow-matching transformer conditioned on the demo predicts dual-arm action trajectories.
- Trained entirely in simulation, it generalizes across folding modes and transfers zero-shot to the real world with no extra data or finetuning.
Paper and code are public; the authors present at Poster Session 2, JW Marriott, Lone Star #51.
More from Embodied
- Must-read papers of the week: HALO humanoid motion, GR00T 2, VLA policies, agent tooling — TheTuringPost · 2026-11-11
- Microsoft's DLSS rival Auto SR launches on Xbox Ally X with handheld 1080p and 120fps support — tomwarren · 2026-11-11
- NVIDIA's PhysicalAI-Human-1M-100K robot dataset trends on Hugging Face — nvidia · 2026-11-11
- Developer runs a hermes-agent lineup on ESP32 with voice-to-voice and tool calling — alexcovo_eth · 2026-11-11
- Tesla surpasses Waymo in Texas robotaxi ops just 8 months after launch — RachelVT42 · 2026-11-11
- Long-horizon mobile manipulation called the standout demo at CoRL — JasonMa2020 · 2026-11-11