Show HN: VLDA-learn compact Deltas as actionable transition interfaces for robot

Category: ai-ml

Tags: training-framework, llm, vlm, embodied-ai, megatron, pytorch, baidu

Score: 8.0/10 (Innovation: 7, Technical: 9, Documentation: 8, Utility: 8)

LoongForge is a production-grade training framework from Baidu's Baige team for LLMs, VLMs, diffusion, and embodied models, built on Megatron-LM and a torch-native backend with claimed up to 4.38x throughput gains while keeping loss aligned. It ships ready-to-run recipes for 40+ models, checkpoint conversion, and data tooling, and is proven at 5,000+ XPU scale, including the new VLDA vision-language-delta-action model with strong LIBERO and real-robot results.

Target audience: ml engineers and researchers training large models; embodied-ai / robotics researchers; infrastructure teams scaling distributed training

Repository: https://baidu-baige.github.io/LoongForge/projects/vlda.html · Python · Apache-2.0 · 705 stars

View on Hacker News