NVIDIA Tutorial: Fine-Tuning Qwen3-VL for Robot Video QA With Cosmos Framework
NVIDIA Developer · youtube · 2026-11-20
NVIDIA published an official tutorial showing how to fine-tune models beyond the Cosmos family using the Cosmos Framework — specifically Qwen3-VL on the RoboVQA dataset for robot video question answering.
Key points:
- The entire pipeline is driven by a coding agent via an agent skill from a single prompt: environment setup, data download/prep, preflight checks, baseline eval, training, export, and post-training scoring
- Evaluation uses held-out robot clips excluded from training, comparing answers before and after fine-tuning
- Demo ran on an NVIDIA GB300 with about two hours of training
- Code and recipes are open-sourced in NVIDIA/cosmos (COOKBOOK.md), plus a technical blog
More from coding & agent
- OpenAI updates prompting guidance for live-1.1: languages and regional accents — pbbakkum · 2026-11-20
- Rippling AI lets any employee prompt their way to custom business tools — andreisavu · 2026-11-20
- Google's Live Interactions API goes live: Genie 4 gets native WebRTC and WebSocket support — _philschmid · 2026-11-20
- Watching an AI agent click through Plotzy's UI: a temporary pattern, says dev — NathanSRobinson · 2026-11-20
- Building an Email Routing Agent with DSPy Weight-Space Programming on 600M/2B Local Models — lateinteraction · 2026-11-20
- Gemini 4 Neon codes a Chamonix ski sim from scratch: ~10k lines, zero assets — flngr · 2026-11-20