openpi

ACOT-VLA-WM is an advanced Vision-Language-Action robotics framework that uses a predictive world model to generate visual subgoals, improving long-horizon robotic manipulation accuracy and robustness.

Open-source implementation of "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"

Real-time inference engine for openpi

Unofficial OpenPI extension experiment to build a more complete OpenPI-style VLA engineering stack: pi0.5 semantics, RTC, pi0.6 RECAP/MEM, and pi0.7-inspired world-model ideas.

Deployment-focused fork of OpenPI for LimX TRON2 manipulation — pi0.5 policy serving, task fine-tuning, TRON2 transforms, and real-robot client examples.

Try OpenPI closed loop simulation / scenes generation / 3DGS model scenes in genie_sim

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

ArcheBase Physical AI 开放电子书：VLA、世界模型、机器人学习、控制、评测与开放数据实践

An end-to-end tour of embodied AI world models: Closed-loop VLA simulation, Real2Edit2Real spatial scene editing, generalization benchmarking, and in-world-model RL self-evolution using GE-Sim 2.0 & OpenPI π₀.₅.

Reproducible AMD ROCm evaluation and full three-camera recording for BiGym, OpenPI JAX, MuJoCo and 3DGS.

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.

Kuavo Scene2 VLA: rosbag-to-LeRobot data pipeline, OpenPI π0.5 LoRA training, and ROS/MuJoCo closed-loop deployment

基于 pi0/OpenPI 的 RM 双臂机器人 MuJoCo 仿真与策略接入项目（由 2025 年 6 月小项目整理，可能存在 bug）。

Physical Intelligence — independent third-party profile of a public API surface, by API Evangelist. Physical Intelligence (often styled "Pi" or "π") is a San Francisco-based research company building general-purpose foundation models for robotics with the stated goal of producing learning algorithms that can control any robot to do any task.

Sanitized engineering notes on AI tools, infrastructure, and robotics.

Model-agnostic robot policy deployment with safety checks, dry-run, audit/replay, metrics and FastAPI.

100-task BEHAVIOR-1K (2026) meta-training pipeline for a pi-0.5 VLA policy on 8x B300. Extends the 50-task BEHAVIOR-Challenge-winning checkpoint to 100 activities.

Enpei Episode1 single-arm LeRobot data collection and OpenPI pi0.5 inference client

Fine-tuned π₀.₅ on 54 own-collected dataset for language-conditioned pick-and-place on a Kinova Gen3 robotic arm. Full pipeline: teleop data collection, LeRobot conversion, LoRA training, and closed-loop deployment with remote inference. A vision-language-action policy that picks up a block and puts it in a bowl.

Closed-loop benchmark of 9 VLA policies (GR00T N1.7, pi0/openpi, LoRA vs full fine-tuning) on BEHAVIOR-1K. Imitation learning / behavioral cloning with diffusion action heads in PyTorch, plus a reproducible eval harness, unattended training automation, and raw metrics.