다시 X(엑스) 활동 열심히 해봐야지
다시 열심히 해보려는 이유...
어이없게도 1년치 프리미엄 결재가 되어버림ㅋㅋㅋ;;
작년에 1년치 결재해서 사용했는데 오늘 자동 결재로 갱신되서 강제 1년 더 사용..;;
그래서 결재된김에 다시 활동 열심히 해서 이번에는 정직원 달성이나 해볼까함
3× FR3 + Gemini Robotics ER + Isaac Sim + ROS 2.
Supervisor graph on the left. Arms on the right.
Concurrent pick, pipelined stack, 9-layer target.
Agents propose the DAG. The runtime owns the motion.
Same loop I want on lab tools: plan → gate → execute.
I wanted to test Jev's spatial ability.
In a loop I gave it the:
- Goal
- Current geometry and contacts
- controls and their predicted effects
- Previous action outcome
I think it's pretty remarkable how it zero shots the task with no vision capability
HOW GPT-6 ASTRA IS USED IN ROBOTICS AND EMBODIED AI
This is not a single system or a unified benchmark. Rather, it’s a map of experiments that have already emerged.
The repository contains 32 cases across four areas:
- robot control in simulation and the real world;
- calling ready‑made robotics policies;
- transferring real movements to simulation;
- creating environments and training via reinforcement learning.
In some examples, Astra plans the robot’s movement, and the low‑level controller then executes the trajectory.
In others, the model itself writes code for Isaac Sim, MuJoCo, or Newton, sets up the environment, and starts the training.
There are also real‑world demonstrations:
- grasping objects;
- inserting a plug;
- working with a keyboard;
- slicing a cucumber;
- drawing with a brush;
- pick‑and‑place.
- adaptation to a new type of robot.
The most interesting part is working with an unfamiliar environment.
In one of the cases, Astra receives visual context and must figure out how to act in a new configuration, without separate training for each scenario.
In some examples, Astra is responsible only for the task and planning, while the movement is performed by a pre‑trained model - for example, VLA or a specialized controller. In other cases, everything happens only in simulation.
The repository’s authors clearly share this.
They also added an unfortunate example: when reconstructing the robotic arm, the mechanics of the tendons turned out to be simplified, and the deformation of the cables was more illustrative than physically accurate.
Awesome-Astra-Embodied-AI shows the direction of development: the model becomes an interface between the human task, the code, the simulation, and the physical robot.
But for now, this is a catalog of demonstrations and experiments, not a proof of production reliability.
GitHub: https://t.co/PuHiytLFu6