pi05

FlashRT is a high-performance realtime inference engine for small-batch, latency-sensitive AI workloads. The flagship integration is production VLA control for Pi0, Pi0.5, GROOT N1.6, and Pi0-FAST. Also support llm e.g, qwen3.6-27B

Real-time inference engine for openpi

Real-time UR5 control with π0.5 and RTC soft prefix guidance, based on openpi.

Learn to reproduce OpenPI pi05 LIBERO_10 LoRA fine-tuning step by step: data splits, cloud training, WSL deployment, and rollout evidence. | 手把手复现 OpenPI pi05 的 LIBERO_10 LoRA 微调：涵盖数据划分、云端训练、WSL 部署与推理视频及日志。

RealMan banana pick-and-place with pi0.5 fine-tuned on 50 demonstrations. Pilot results: Expert-only 9/10, LoRA 0/8.

Human-in-the-loop teleoperation and data collection for AGIBOT G2A, with VLA deployment, trajectory rollback, and auditable datasets.

Research project on test-time action selection for VLA policies — manuscript under review; code, results & paper will be released after publication.

Fork of openpi: π0.5 fine-tuned with B-spline action targets (B-spline Policy) and executed asynchronously with phase skipping. Includes a deterministic, paired random-latency LIBERO benchmark (Normal 300±60 ms, 2,000 episodes per mode) against async chunking and Real-Time Chunking.

A simple deployment engine for VLA / WAM models.

π0.5 single-arm robotics: dual-camera demonstrations, full fine-tuning on H20, cloud policy serving and real-robot pick-and-place.

Windows-native RM65 data collection, LeRobot v3 publication, and supervised pi0.5 rollout runtime.

Agentic VLA demo: MiniMax LLM agent orchestrates Pi0.5 for long-horizon manipulation in LIBERO sim

π0.5 VLA fine-tuning and real-robot deployment on a dual-arm Kuavo 5 humanoid: LoRA rank requirements measured by SVD spectral energy, a silent non-persistent-buffer loader bug, and honest ablations. Notes and reproducible analyses; no data or weights.

Edge-Oriented Multimodal Inference Acceleration and Benchmarking