SynthICL: synthetic-data-only training hits 79% success on 16 real robot tasks
Ed__Johns · x · 2026-11-12
Edward Johns' group at Imperial College presents SynthICL at CoRL 2026: a framework training in-context imitation learning policies entirely from synthetic RGB data. At test time it adapts to unseen manipulation tasks from a single demonstration without fine-tuning.
Key points:
- Data pipeline generates high-fidelity synthetic demonstrations at scale, avoiding depth sensing, precise camera calibration, and real-world training data.
- Uses a flow-matching transformer policy with subgoal image prediction for more precise, visually grounded control.
- Achieves 79% average success across 16 unseen real-world tasks with one demo, outperforming prior methods.
- Demonstrates sim-to-real deployment.
More from Embodied
- Berkeley open-sources RoboSimStudio: AI agents program industrial robots without training data — berkeley_ai · 2026-11-12
- NXP ships pre-tested drone platform: initial flight in 30 minutes out of the box — shashib · 2026-11-12
- Open-source dual-arm robot solves tangram puzzles with 90 demos, ~60% success — DanGrover · 2026-11-12
- Robotics has three regional playbooks: China volume, Europe inspection, US pilots — ingliguori · 2026-11-12
- 3,500 Laser-Armed AI Submarine Robots Patrol Norwegian Salmon Farms — philfung · 2026-11-12
- Model Autonomously Designs a Surgical-Robot-Style Robotic Arm CAD After 75 Minutes — _Stocko_ · 2026-11-12