Breaking

NVIDIA Shows Cosmos 3 Physical AI Model Post-Trained in Under a Day Using Automated Agent Skills

July 16, 2026 at 12:07 EDT

Foundation Models
AI Agents
Robotics

NVIDIA has demonstrated that its Cosmos 3 physical AI model can be fine-tuned for a specialized real-world task in less than a day, using coding agents to automate the machine-learning workflow and pushing accuracy from 54.41% to 93.35% with minimal human intervention.

July 14, 2026 · NVIDIA

Cosmos 3, Post-Trained in Under a Day

Using coding agents to automate the machine-learning workflow, NVIDIA lifted a physical-AI model from a 54% zero-shot baseline to 93% accuracy on a real traffic-safety benchmark — with minimal human intervention.

93.35%

Peak accuracy after AutoML sweep

< 1 day

Total post-training time, mostly automated

7×

Fewer GPU hours vs. full fine-tuning (LoRA)

Accuracy climb on the Woven Traffic Safety benchmark

54.41%

Zero-shot
baseline

→

87.14%

Single LoRA run
(~30 min)

→

93.35%

AutoML sweep
(43 trials)

The agent-driven workflow

Coding agent

Codex / Claude Code / Cursor absorbs the MLOps glue code

→

LoRA fine-tune

One 30-min run on Cosmos 3 Nano (~16B params)

→

AutoML sweep

43 trials · 19.5 hrs · 8 A100 GPUs in parallel

→

Serve via NIM

LoRA adapter as an OpenAI-compatible endpoint

Why it matters

Targets the "last-mile accuracy" problem in physical AI — closing the gap on domain-specific camera angles and edge cases, with 56 open-source TAO skills and a real Toyota Woven dataset of 8,000+ samples.

Still open

Published only recently, so large-scale independent evaluation is limited. How the approach generalizes across other domains and edge cases will require broader testing.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.