Robot video labelling: Nvidia's Flora costs 72 cents an hour

What happened

Nvidia has open-sourced Flora, a data engine that labels one hour of robot video for 72 cents, against $88 for the same work using GPT-6 Astra, according to reporting from The Decoder. Flora also scored slightly higher than GPT-6 Astra on labelling quality.

The system is built on three fine-tuned Qwen3.5 models. It detects errors in robot actions more reliably than the competing systems covered in the reporting, but it sometimes falls short on descriptive accuracy, meaning the written description of what a robot is doing may be less precise even when the error detection is stronger.

Nvidia is positioning Flora as one component of a wider strategy to control the entire robotics value chain. The headline figure of 125 times cheaper is the outlet's own framing; the per-hour prices reported ($88 against 72 cents) work out to roughly 122 times.

Why it matters

Labelling video is one of the least visible constraints in building capable robots. Models that control physical machines need large volumes of annotated footage showing what an action was, whether it succeeded and where it went wrong. When that labelling is expensive or inconsistent, teams either limit their datasets or accept weaker training. A tool that is both far cheaper and better at catching action errors lowers that barrier, and an open-source release means organisations can run the pipeline on their own infrastructure rather than paying per call to a closed service.

For leaders in digital transformation, the more important shift is operational. Data preparation moves from an outsourced or ad hoc task to a repeatable internal capability. The weakness reported in descriptive accuracy also matters: it suggests that automated labelling still needs human review wherever the explanation of an event, rather than the detection of a fault, is what downstream teams rely on.

By the numbers

• 72 cents to label one hour of robot video with Flora, compared with $88 using GPT-6 Astra.
• 125 times cheaper, the figure used in the headline reporting (the per-hour prices imply roughly 122 times).
• Three fine-tuned Qwen3.5 models form the basis of the open system.
• One hour of robot video is the unit of comparison used in the reported cost test.

The Renascence take

Most coverage will focus on the price gap. The more useful question for operators is where the errors go, because a labelling engine that is good at spotting failures but weaker at describing them shapes what teams can trust.

The cheapest annotation is not the best annotation; the one that matters is the one that tells your team what went wrong and why. Customer-obsessed operators should treat Flora as a way to label far more footage, then route its descriptions through human review before they inform decisions about customers or service. Use the cost saving to fund that review loop rather than to cut it. Experience quality in robotics is decided at the moment a person relies on a machine's account of its own behaviour, and that moment deserves the same scrutiny as the detection rate.