Qwen-RobotManip: alignment before scale for robotic manipulation foundation models
rsasaki0109 · x · 2026-09-03
The paper introduces Qwen-RobotManip, a generalizable vision-language-action foundation model built on Qwen-VL / Qwen3.5-4B. It pairs a vision-language backbone with a flow-matching Diffusion Transformer action expert for continuous action generation while preserving perception and language grounding.
The central principle is alignment before scale: robot manipulation data is inherently heterogeneous across embodiments, action spaces, cameras, coordinate frames and task distributions. The work proposes a unified alignment framework across representation, motion and behavior so multi-source training becomes coherent, trained using only open-source and egocentric datasets.
More from Embodied
- Sentdex trains robots to go from lying down to a half kneel, eyeing a full stand — Sentdex · 2026-10-21
- NVIDIA open-sources MotionBricks: 350K motion skills at 15,000 FPS, 2ms latency — aftahi_ai · 2026-10-21
- NVIDIA Cosmos3-H-Dreams: Generative Simulation Brings Clinical Surgery to Life — Hugging Face Blog · 2026-10-21
- FLoTS-SLAM uses optical-flow residuals to hit 0.73 DM-IoU in dynamic monocular SLAM — zhenjun_zhao · 2026-10-21
- Plato ring launches at $199 to turn near-silent speech into AI commands — TinfoilTricorn · 2026-10-21
- Dream-1 controlled 31 unseen robot types across 53 countries within 24 hours — gharik · 2026-10-21