Full-Stack-Entity/Robotwin-demo-clean-50-artifacts
收藏资源简介:
这个仓库存放的是一次RoboTwin单任务闭环实验的关键结果产物,而不是官方完整数据集。实验完整跑通了从任务程序生成专家轨迹到策略重新部署并评测的整个链路。关键结果显示,在同分布(demo_clean -> demo_clean)下的成功率为0.64,而在跨配置(demo_clean -> demo_randomized)下的成功率为0.0。实验体现了RoboTwin作为具身智能实验工程流水线的价值,包括环境层、专家层、数据层、训练层、部署层和评测层的严格流程。
This repository contains the key results of a RoboTwin single-task closed-loop experiment, not the official complete dataset. The experiment fully executed the entire pipeline from generating expert trajectories with the official task program to redeploying the policy and evaluating it. Key results show a success rate of 0.64 in the same distribution (demo_clean -> demo_clean) and 0.0 in the cross-configuration (demo_clean -> demo_randomized). The experiment highlights the value of RoboTwin as a rigorous engineering pipeline for embodied AI experiments, including strict processes at the environment, expert, data, training, deployment, and evaluation layers.




