zjwzcx/Awesome-Astra-Embodied-AI

Awesome Astra Embodied AI

GPT-6 Astra with embodied AI / robotics workflows and demos.

Contents

Cases are grouped by workflow and ordered by publication date, newest first within each section.

• 🤖 Zero-shot Control
• 🧪 Deploy in Simulation — 12 cases
• 🌍 Deploy in Real World — 10 cases
• 🧠 Agentic Policy Calls — 1 case
• 🔄 Real-to-sim Replay / Data Rollout — 6 cases
• 🛠️ Astra Builds RL Training Environments and Training — 6 cases
• 📊 Public Benchmark & Evaluation Reports

🤖 Zero-shot Control

🧪 Deploy in Simulation

Astra directly performs zero-shot control in a simulator.

Case 1: Dual-ALOHA Spatial-constraint Puzzles

Source / Credit: Qineng Wang (@qineng_wang), X demo of Astra planning two Dual-ALOHA spatial-constraint tasks, with an accompanying interactive replay and methods page.

Published: 2026-09-15

Astra plans motions that separate interlocked parts and thread a rope through three rings. The claw task is a kinematic replay from a pregrasped state; the rope task uses native MuJoCo dynamics, while both use disclosed ideal-grasp assumptions and demonstrate specific planned motions rather than an online policy.

Case 2: Vision-only Humanoid Body Control

Source / Credit: ZQ, Rednote demo where Astra observes a simulated humanoid through the robot camera and controls its body without privileged state information.

Published: 2026-09-13

Astra uses camera observations to control the humanoid and adjusts its actions during execution without access to privileged simulator state.

Case 3: G1 Cola Bottle Pick-up in Isaac Sim

Source / Credit: Flood Sung (@RotekSong), X demo of the GPT-6 Codex agent controlling a Unitree G1 in Isaac Sim.

Published: 2026-09-13

Astra performs high-level planning for the cola-bottle pick-up, and the GEAR-SONIC planner converts the plan into a whole-body qpos trajectory for the simulated G1 to execute.

Case 4: Quadruped Task via Five-key-joint Trajectory

Source / Credit: Akira Sasaki (@gclue_akira), X quadruped locomotion demo where Astra outputs five key-joint trajectories for MuJoCo execution.

Published: 2026-09-11

Astra supplies the sparse trajectory while the low-level controller executes the quadruped behavior in MuJoCo.

Case 5: G1 Navigation Trajectory Tracked by SONIC

Source / Credit: Flood Sung (@RotekSong), X demo of Astra producing a navigation trajectory for SONIC to track.

Published: 2026-09-11

Astra performs high-level navigation planning, and the GEAR-SONIC planner converts the plan into a whole-body qpos trajectory for the simulated G1 to execute.

Case 6: Robot Hands Solve a Rubik’s Cube

Source / Credit: Ze Yanjie (@ZeYanjie), X demo titled “GPT6 Astra solved Rubik’s Cube with robot hands.”

Published: 2026-09-10

Astra performs zero-shot dexterous cube manipulation in simulation.

Case 7: Dexterous Apple-stem Grasp in SuperDex

Source / Credit: Kiki Huang, Rednote demo using the Meta SuperDex physics backend and MuJoCo rendering.

Published: 2026-09-10

Astra controls a dexterous hand to grasp the narrow stem of an apple in simulation.

Case 8: GPT-6 Astra on HumanCLAW-Bench

Source / Credit: Jiawei Gu (@Kuvvius), X benchmark demo using the open-source HumanCLAW-Bench harness and motion generator.

Published: 2026-09-10

Astra completes benchmark navigation and interaction tasks through the simulation harness.

Case 9: Whole-body Trajectory with a Whole-body Controller

Source / Credit: 橘子不是唯一的水果, Rednote demo where Astra generates a whole-body trajectory and a whole-body controller executes it.

Published: 2026-09-08

Astra generates the trajectory while the whole-body controller closes the execution loop.

Case 10: Isaac Sim Cube Grasp from One Prompt

Source / Credit: 神秘小孙, Rednote demo of Astra building a depth-camera robot-arm cube-grasp scene in Isaac Sim.

Published: 2026-09-08

Astra receives one task sentence and produces a working simulated grasp demo.

Case 11: Physically Writing a Fibonacci Sequence

Source / Credit: Dmytro Hrybov (@dimentary), X demo of Astra generating a Unitree G1 motion for a physical Fibonacci-writing task.

Published: 2026-09-08

Astra attempts a long-horizon simulated manipulation task that ends with generated code and robot motion.

Case 12: LLM Harness for Tabletop Robot Control

Source / Credit: Jiafei Duan (@DJiafei), X demo of a GPT-6 Astra harness solving a tabletop grasp-and-place task.

Published: 2026-09-06

Astra writes and runs the control loop for a simulated tabletop manipulation task.

🌍 Deploy in Real World

Real-robot demonstrations with explicit hardware and deployment context.

Case 13: Learning to Type and Self-express on a Keyboard

Source / Credit: Kaifeng Zhang (@kaiwynd), X demo where Astra learns to operate a keyboard with a real robot after being asked to express itself.

Published: 2026-09-12

Astra turns an open-ended semantic instruction into physical key presses, learning the robot control through visual feedback in about 40 minutes.

Case 14: Marker Grasp from Low-level Control Only

Source / Credit: star大小变, Rednote demo where Astra discovers joint-to-end-effector control and grasps a marker without prior skills.

Published: 2026-09-12

Astra explores the robot’s joint readings and reaches a marker grasp in about 30 minutes.

Case 15: GPT-Policy-Eval One-shot Plug Insertion

Source / Credit: GPT-Policy-Eval, one-shot video demonstration of GPT-6 Astra guiding a real robot to grasp and insert a plug with visual feedback.

Published: 2026-09-11

Astra executes a contact-rich plug-insertion task from a single video demonstration without VLA, RL, or DAgger; see the complementary RoboCurve evaluation for controlled robot-arm results.

Case 16: Mobile Manipulation with In-context Learning

Source / Credit: Axel (@ax_pey), X demo explicitly showing GPT-6 Astra in-context learning across environments, viewpoints, and layouts.

Published: 2026-09-11

Astra infers mobile-manipulation behavior from visual context without a task-specific text prompt.

Case 17: Rapid Adaptation to an Unseen Robot Embodiment

Source / Credit: Lucas Cassiano (@lucascassiano), X demo giving Astra full access to the robot hardware and the Vitrus AI robotics OS.

Published: 2026-09-09

Astra learns to control a robot embodiment it has not seen before, without human egocentric data or a VLA, showing rapid online adaptation to a new control interface.

Case 18: ENPIRE Robotics Harness In-context Learning

Source / Credit: Tonghe Zhang (@TongheZhang01), X demo of GPT-6 Astra performing robot in-context learning through the ENPIRE harness.

Published: 2026-09-09

Astra learns the real-world behavior from demonstrations in context, without task-specific retraining.

Case 19: Cucumber Slicing with Loop-ROS

Source / Credit: 盒子桥, Rednote demo of GPT-6 Astra directly controlling a robot through Loop-ROS.

Published: 2026-09-09

Astra controls the robot arm through the slicing sequence in the real world.

Case 20: Painting the Golden Gate Bridge from a Semantic Prompt

Source / Credit: thijs (@cdngdev), X demo where Astra is given a robot, a paintbrush, and a camera, then asked to paint the Golden Gate Bridge in the real world.

Published: 2026-09-08

Astra translates the high-level visual concept into physical brush strokes and progressively improves the painting across attempts using camera feedback.

Case 21: Direct End-effector Pose Control

Source / Credit: Loule, Rednote demo of Astra directly outputting end-effector poses for a robot arm.

Published: 2026-09-08

Astra places the longest piece of bread into a basket using third-person and wrist cameras.

Case 22: Piper Carrot Pick-and-place

Source / Credit: 虽然不但是, Rednote demo using only Codex/GPT-6, Piper, and the RealSense SDK.

Published: 2026-09-05

Astra directly controls Piper through repeated visual pick-and-place attempts.

🧠 Agentic Policy Calls

Astra performs high-level task understanding and decomposition, then calls a pretrained embodied foundation model for low-level control.

Case 23: Zero-shot Task Execution through FluxVLA

Source / Credit: Jikun, Rednote demo pairing GPT-6 Astra with the pretrained FluxVLA policy.

Published: 2026-09-09

“Zero-shot” refers to Astra’s task inference and planning; FluxVLA executes the low-level embodied actions.

🔄 Real-to-sim Replay / Data Rollout

Workflows that reconstruct or replay real-world trajectories, demonstrations, and environments in simulation.

Case 24: Lab Kitchen Reconstruction with Articulated Objects

Source / Credit: Frank ZY Dou, Rednote demo reconstructing a lab kitchen from a 20-second monocular RGB video with movable cabinets and other articulated structures.

Published: 2026-09-12

Astra iteratively models a real kitchen and its articulated objects into a simulation scene.

Case 25: Rope-driven Dexterous Hand Reconstruction

Source / Credit: Dmytro Hrybov (@dimentary), X attempt to recreate 1X's tendon-driven hand demo from a reference video in MuJoCo.

Published: 2026-09-10

Astra rebuilds the hand and motion, but uses simplified mechanics and illustrative cable deformation rather than full tendon-transmission physics.

Case 26: Tendon-driven Dexterous Hand Motion Reconstruction

Source / Credit: Jake Fitzgerald (@earthtojake), X demo in which Astra designs and reconstructs the motion of a tendon-driven robot hand.

Published: 2026-09-09

Astra reconstructs the tendon-driven hand's motion in a simulated model, including cable-actuated finger movement.

Case 27: Dexterous Hand-object Data Rollout

Source / Credit: Lingxiao (@Lingxiao234), X demo showing two videos driving real-to-sim reconstruction and physical retargeting to Wuji hands.

Published: 2026-09-09

Astra reconstructs hand-object interaction from videos without explicit states or actions.

Case 28: Video in → Physics out

Source / Credit: xiao hu (@huxiao93612565), X demo of Astra writing hand tracking, IK retargeting, and grasp-refinement code for a 44-DOF hand.

Published: 2026-09-09

Astra turns visual hand motion into a physics-backed dexterous-hand replay.

Case 29: Multi-view Real-to-sim Reconstruction

Source / Credit: Lingxiao (@Lingxiao234), X demo combining multi-view RGB, robot actions, camera calibration, assets, system identification, MuJoCo, and Blender.

Published: 2026-09-07

Astra builds a replayable simulator from demonstrations, geometry, and physical parameters.

🛠️ Astra Builds RL Training Environments and Training

Workflows where Astra helps create environments, task definitions, training code, and experiment iterations.

Case 30: Sharpa Hand Pen-spinning RL in Isaac Lab

Source / Credit: Wentao Zhu (@walterzhu8), X report of an autonomous Astra run by his student Chengyang Li to build and train a dexterous-hand pen-spinning task.

Published: 2026-09-16

Astra creates the pen mesh, implements the Sharpa-hand task in Isaac Lab, trains the PPO policy, and produces a visualization video during an autonomous run of about a day and a half.

Case 31: RL-trained Duck Robot Demo

Source / Credit: 拂晓时分_茉莉飘香, Rednote demo reporting an RL-trained duck robot generated from one image and one description.

Published: 2026-09-11

Astra builds and trains a locomotion demo from a compact visual specification.

Case 32: Quadruped Locomotion System from RL

Source / Credit: Akira Sasaki (@gclue_akira), X report of Astra designing a robot dog, iterating 25 loops in five days, and training nine motions with RL.

Published: 2026-09-11

Astra co-designs the quadruped and trains a simulated locomotion system; real-hardware debugging is planned rather than completed.

Case 33: Dexterous In-hand Manipulation RL

Source / Credit: 十一, Rednote RL demo of a dexterous hand manipulating a walnut; self-collision was not enabled in the reported run.

Published: 2026-09-11

Astra trains the in-hand manipulation behavior while exposing the reported collision-model limitation.

Case 34: Office Scan → Newton / G1 Humanoid Gym

Source / Credit: Jiarui Xu (@Jiarui_X), X demo where Astra rebuilds an office scan in Blender, exports USD, and creates a G1 walking scene in Newton.

Published: 2026-09-11

Astra creates the environment and training-ready humanoid simulation scene.

Case 35: Isaac Sim Environment, PPO Training, and Tuning

Source / Credit: 十一, Rednote demo of Astra building an Isaac Sim RL environment, configuring PPO, and tuning the run.

Published: 2026-09-10

Astra handles environment construction, training configuration, and iteration in one workflow.

⚠️ Failure Cases

These entries index representative partial or unsuccessful outcomes. A failure can reflect modeling, calibration, embodiment, or tool limits in one setup and does not prove that Astra lacks the underlying capability.

• Rope-driven Dexterous Hand Reconstruction: partial real-to-sim reconstruction with simplified tendon mechanics and illustrative cable deformation; full tendon-transmission physics remains unfinished.

📊 Public Benchmark & Evaluation Reports

Public evaluation repositories and result reports for GPT-6 Astra in embodied-AI settings. Results below are reported by the respective maintainers and should be interpreted within each project's disclosed hardware, task, and evaluation protocol.

• RoboCurve GPT-6 Astra evaluation: controlled YAM-arm comparison reporting 19/20 bowl-task completions for Astra and 80% fewer output tokens; published 2026-09-04.
• GPT-Policy-Eval: open evaluation repository reporting one-shot transfer from a video demonstration to a real-robot, contact-rich plug-insertion task, wi