openvla

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

Robotics Data Toolkit | Convert between robotics dataset formats (RLDS, LeRobot v2/v3, Zarr, HDF5, Rosbag). Inspect, visualize, and analyze datasets. Works with HuggingFace Hub. Built for OpenVLA, Octo, LeRobot, and Diffusion Policy workflows.

OpenVLA Lightweight Version(0.5B). It uses qwen2-0.5B and fine-tunes using mllm format, without occupying LLM's inherent tokens. It represents actions with an additional 271 tokens.

PickAgent: OpenVLA-powered Pick and Place Agent | Gradio&Simulation | Vision Language Action Model

ROS2-native runtime, benchmark, and adapter hub for Vision-Language-Action models.

Train and deploy Vision-Language-Action models natively on Apple Silicon.

Independent VLA research notes: OpenVLA / π0 / Spirit paper reviews, LIBERO reproduction, XLeRobot integration. Transitioning from AD motion planning to embodied AI.

Think Less, Act Early: Reinforced Latent Reasoning with Early Exit in Vision-Language-Action Models

Using VLM-based visual question answering to perceive scenes and control robots in MuJoCo simulation.

FR3 robot in Gazebo integrated with 4-bit quantized OpenVLA and MoveIt

A full-stack Embodied AI simulation suite powered by Genesis World: featuring OpenVLA closed-loop evaluation, procedural scene generation, massively parallel GPU RL (Unitree Go2), and multi-physics coupling (PBD/SPH).

OpenVLA-based grasping in a custom PyBullet environment

A RoboNix Skill for experience-memory retrieval and verified action reuse across OpenVLA and π0.

Safety monitors for learned robot policies under distribution shift. A benchmark for which monitor still works once the deployment distribution moves.

Do robot policies actually use their cameras? Reproducing ACT, OpenVLA, SmolVLA, Diffusion Policy, π0 and π0.5 in simulation, plus vision-dependence experiments (blackout, perturbation, trajectories, state-zeroing fine-tune).

Reproducible audit of SAFE VLA failure detection on OpenVLA/LIBERO — pooled AUC inflates +0.09 over macro within-task; functional CP catches only 7-9% at ≤5% FPR

Unofficial PyTorch reproduction for OpenVLA: An Open-Source Vision-Language-Action Model.

Simulated UR5e + MoveIt 2 cell for testing vision-language-action policies in ROS 2 Jazzy. Swap SmolVLA, OpenVLA-7B (4-bit) or GR00T N1.7 behind one ZeroMQ protocol — all runnable on a 6 GB GPU.

VLA manipulation ablations, held-out generalization, and an OpenVLA-7B feasibility study, all measured on a 6GB laptop GPU

π0.5 hears the instruction. What it hears does not decide what it does. Paired-prompt probing and causal intervention on OpenVLA and π0.5 (LIBERO), with an interactive explorer.