smolvla

仅需Python基础，从0构建自己的具身智能机器人；从0逐步构建VLA/OpenVLA/SmolVLA/Pi0， 深入理解具身智能

FastCrest Tether: the OSS edge-to-cloud AI deploy CLI. Optimize, verify, deploy across Jetson, RTX, Apple Silicon, AMD. Hybrid edge-cloud inference with parity certs.

实机部署so101流程和踩坑全记录，使用lerobot框架so101真机。在本地 ubuntu22+5070显卡 搭建环境，部署，录数据，训练smolvla，推理全流程。

This repository implements pi0, pi0fast and smolvla models on sim and real robots using Lerobot interface

Automated agentic research loop for vision-language-action (VLA) robot policy self-improvement

Provael — red-team open Vision-Language-Action (VLA) robot policies in simulation and report an Attack Success Rate (ASR). Prove it. Prevail.

End-to-end SO-101 tic-tac-toe: printable hardware, local UI, LeRobot dataset, SmolVLA policy and guarded Strands agent

Vision-language-action stack for contact-rich robot manipulation on sub-$300 hardware. SO-101 + LeRobot; ACT 18/20 and SmolVLA 16/20, both generalizing to a held-out workspace cell.

Explore recorded Physical AI: Genesis × Newton GPU timesteps, Microduck joints, SmolVLA failures and cloth. Inspect source data, offline replays and editable OpenUSD.

Manual real‑world benchmarking of ACT/SmolVLA on a physical SO101 robot using LeRobot.

Agentic planning for Vision-Language-Action robotics — orchestrating SmolVLA/LeRobot policies with LangGraph for multi-step robot manipulation.

在 RTX 3060 6GB 上复现 SmolVLA × LIBERO，并开展低显存 LoRA 微调、多随机种子评测与遗忘控制实验。Low-VRAM SmolVLA × LIBERO reproduction with LoRA adaptation, fixed multi-seed evaluation, and forgetting controls.

Unified pipeline for collecting, labeling, converting, and training robot demonstration data with SmolVLA, Pi0, Pi0.5, or ACT policies - extended with Qwen3-VL offline phase annotation, a runtime FSM, and three phase-conditioned policy variants.

Per-Group Error, Not Total MSE. Fine-tuning VLAs on the Toyota HSR. ICRA 2026 Workshop From Data to Decisions, 1st place out of 36.

Reproducing SmolVLA-450M on LIBERO benchmark with cross-embodiment experiments | LeRobot

A heterogeneous dual-model robot navigation system pairing Falcon Mamba (SSM) as a fast reflex layer with SmolVLA (VLA) as a deliberate planning layer, arbitrated by a priority controller, fully simulatable in Wokwi on ESP32. as a part of learning project.

Reproducible dual-camera SO-101 SmolVLA training, RTC deployment, and dual-instruction hardware evaluation with LeRobot

How much does it cost to teach a robot a new part? Measuring the demonstration-count vs. success-rate curve for SmolVLA on LIBERO, on a single 8GB consumer GPU.

Terminal control center for the LeKiwi robot workflow, built with LeRobot and pyratatui.