Lightning Talk: Full-Stack PyTorch Robotics VLA: From Data To...
summary
A PyTorch-native pipeline that takes teleoperation data through training to optimized edge inference, claiming an end-to-end Physical AI flow that runs on CPU/GPU/NPU with operator fallback to native PyTorch. Implements PyTorch Lightning for distributed mixed-precision training, export via ExecuTorch/OpenVINO with OpenVINO Quantizer (NCF) and AOT compilation, and a small-memory runtime. Demonstrates on a WidowX pick-and-sort case validating latency and numerical tolerances, and exposes GUI/CLI/API for data capture, fine-tuning, export and on-device inference as Full-Stack PyTorch Robotics VLA.
AI-written summary · about
chapters (10)
- 0:00Full-stack PyTorch robotics VLA
- 1:04Robotics AI lifecycle
- 2:10GUI data collection & training
- 3:35CLI/API for policy training
- 4:41High-level PyTorch architecture
- 6:50Optimizing & exporting models
- 7:12Inference with OpenVINO
- 9:19Executor Torch for hardware
- 12:58Edge model export
- 13:41Robot inference
related talks
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...
11mLightning Talk: Accelerating PyTorch Performance with OpenVINO
30mBreaking Heterogeneity Barriers: Unified Cloud-to-Robot AI System SW Stack for...
11mSponsored Session: Lightning Talk: PyTorch and Democratization of AI Accelerators
29mMulti-Accelerator PyTorch Serving With NxD Inference and vLLM
23mAn Open Source Post-Training Stack: Kubernetes + Ray + PyTorch + vLLM
08mSponsored Session: Lightning Talk: PyTorch in Production: Boosting LLM Training and...
21mImplementing your AI strategy with PyTorch Lightning in the cloud