# Hugging Face’s JAM model cuts robotics training costs by 73% without video generation

Digest AI · Robotics & Physical AI · published 2026-10-27T21:58:00Z

Canonical: https://digestai.news/story/hugging-faces-jam-model-cuts-robotics-training-costs-by-73-without-vid

## Summary

Hugging Face released a paper titled *Vision-Language-Action or World-Action? Just Action Models*, proposing a new approach called **JAM** that eliminates video generation during robotics training. The method uses a pretrained visual backbone and an action expert, trained via a *foresight loss* that predicts frozen future frame encodings—without retaining the generation branch at inference. 65.27%). The key innovation is a **73% reduction in FLOPs per sample** during training, measured on 8 NVIDIA GH200 GPUs, by caching foresight targets.

The paper argues the debate over vision-language-action (VLA) vs. world-action models (WAM) is overstated, as the performance gap isn’t tied to a single component—backbone initialization, future-observation objectives, or generation machinery. Cross-backbone tests on LIBERO-Plus show comparable results across Qwen3.5-2B (81.14%), Ovis-U1 (79.95%), and DINOv2-L (70.39%), all under JAM’s foresight recipe. The authors release public code built on FastWAM and ImageWAM, positioning JAM as a compute-efficient alternative to generative approaches in robotics.

## Key points

- JAM cuts robotics training FLOPs by 73% vs. FastWAM by caching foresight targets on 8 GH200 GPUs
- On LIBERO benchmark, JAM improves success rates from 97.0%/55.3%/91.2% to 98.2%/64.5%/91.4%
- Paper claims vision-language backbones and generative models perform equally well under JAM’s foresight loss

## Why it matters

JAM could redefine robotics training efficiency by removing video generation, slashing compute costs while matching or exceeding generative models’ performance. This challenges industry assumptions about pretrained vision-language models’ necessity and may accelerate deployment of AI-driven robots.

## Sources

1. [Hugging Face Paper: Robot Policies Don't Need Video Generation](https://aiweekly.co/alerts/hugging-face-paper-robot-policies-dont-need-video-generation) (aiweekly.co, 2026-10-27)

Part of the developing story: [InclusionAI's Accelerating AI Model Saga](https://digestai.news/thread/inclusionai-releases-ming-3-0-audio-flash-a-multilingual-audio-language-model) (4 stories)

## Cite

Digest AI, "Hugging Face’s JAM model cuts robotics training costs by 73% without video generation", 27 October 2026, https://digestai.news/story/hugging-faces-jam-model-cuts-robotics-training-costs-by-73-without-vid

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/hugging-faces-jam-model-cuts-robotics-training-costs-by-73-without-vid.json