Embodied AI Sprint
Self-imposed eight-week sprint toward a summer-2026 research internship in embodied AI. The stack pairs a π₀ VLA model with Isaac Lab for sim, LeRobot for data and training, and ROS 2 on an SO-101 arm for real-robot work.
Milestone M0 — environment. Done. Isaac Sim 4.5.0, Isaac Lab 2.x, PyTorch 2.5.1 on CUDA 12.4, running on a workstation with an RTX 4070 Ti Super (16GB). Apptainer containers for HPC portability.
Milestone M1 — Franka Reach RL. In progress. Baseline PPO / SAC on Isaac Lab’s Franka-Reach task as a sanity check before touching manipulation.
Later milestones. π₀ fine-tune on LeRobot datasets · sim-to-real on the SO-101 · a minimal language-conditioned pick task.
Code and weekly logs at github.com/LUOSYrrrr/EMBODIED-AI.
为 2026 年夏季具身智能研究实习准备的八周冲刺。技术栈：π₀ VLA 模型 做策略，Isaac Lab 做仿真，LeRobot 做数据和训练，真机用 ROS 2 驱动 SO-101 机械臂。
M0 —— 环境。 已完成。Isaac Sim 4.5.0、Isaac Lab 2.x、PyTorch 2.5.1 + CUDA 12.4，跑在 RTX 4070 Ti Super 16GB 工作站上；用 Apptainer 容器保证能迁到 HPC。
M1 —— Franka Reach RL。 进行中。在 Isaac Lab 的 Franka-Reach 任务上跑 PPO / SAC 基线，作为接入 manipulation 前的 sanity check。
后续里程碑。 在 LeRobot 数据集上 fine-tune π₀ · SO-101 上做 sim-to-real · 一个最小的语言条件抓取任务。
代码和每周进度记录见 github.com/LUOSYrrrr/EMBODIED-AI。