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

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 — 4 cases
• 🔄 Real-to-sim Replay / Data Rollout — 7 cases
• 🛠️ Astra Builds RL Training Environments and Training — 6 cases
• 📊 Public Benchmark & Evaluation Reports

Astra directly performs zero-shot control in a simulator.

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.

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.

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.

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.

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.

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.

Source / Credit: Kiki Huang, Rednote demo using the Meta SuperDex physics backend and MuJoCo rendering. The demo package is now open source in DexLab.

Published: 2026-09-10

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

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.

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.

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.

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.

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.

Real-robot demonstrations with explicit hardware and deployment context.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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

Source / Credit: Seungyeon Kim et al., Recursive Harness Distillation (RHD). Project Page, Paper, Hugging Face, and LinkedIn.

Published: 2026-09-27 (arXiv submission)

Astra distills its robotic problem-solving experience into a playbook and recursively refines it through Luna’s execution feedback, enabling the cheaper Luna agent to guide a frozen VLA.

Source / Credit: Bingao Chen, Haoquan Fang, and C. Karen Liu (Stanford University) — Blog, Videos, and Code.

Published: 2026-09-22 (project website)

GPT-6 Astra plans grounded subtasks from task instructions and visual history, a fine-tuned π0.5 VLA executes them, and a fine-tuned Qwen3-VL-4B model monitors subtask completion. Completion events trigger new Astra planning calls, while the VLA continues executing the current subtask between calls. This three-tier system separates frequent progress checks from high-level reasoning for memory-dependent manipulation in simulation; see the evaluation report for results on RoboMME.

Source / Credit: Harness VLA / RPent team — Project, Video, Paper, and Code (RPent).

Published: 2026-09-17 (video added to the project page)

GPT-6 Astra acts as the planner in RPent, combining memory and visual feedback to compose frozen VLA calls with a fixed library of analytic motion primitives. The real-robot demonstration shows task retargeting, ordered execution, sorting, and recovery from failed grasps without fine-tuning the VLA. The sequence above follows one plate-sorting task from a failed grasp through retry and placement.

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.

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

Source / Credit: watchtower, Rednote demo of a full GPT-6 Astra real-to-sim-to-real loop on a humanoid robot in an office.

Published: 2026-09-23

Astra rebuilds the real office as a simulation scene (real2sim), estimates contact points and other key supervision signals, and trains a reinforcement-learning manipulation policy from them. In the reconstructed scene it decides where to grasp and where to push while SONIC drives the humanoid's whole-body motion, and the trained policy is then deployed back onto the real robot (sim2real), which pushes a cart through the doorway. The posted video runs at real speed in a single unedited take, and longer-horizon tasks are planned.

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.

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.

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.

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.

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.

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.

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

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.

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.

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.

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.

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.

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.

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 evaluation repositories and result reports for GPT-6 Astra in embodied-AI settings. Entries are ordered by publication date, earliest first, with the official RoboCurve evaluation listed first. 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, without VLA, RL, or DAgger; published 2026-09-11.
• GPT-as-Policy, from Galbot (银河通用): public benchmark repository and result report evaluating GPT-6 Astra as a robot policy; published 2026-09-16.
• RoboDojo GPT-6 Astra evaluation: zero-shot evaluation through the fixed, non-learned RoboProbe harness on 42 simulation tasks (2,100 trials, one seed), reporting 28.97 Average Score and 22.48% Average SR; real-robot results are diagnostic only after testing was halted for safety; published 2026-09-16.
• RPent (Recursive Physical Agent), from the RLinf team: open recursive-agent harness that pairs an agentic planner (Codex / GPT-6 Astra) with frozen VLA primitives, with a public leaderboard covering LIBERO, LIBERO-PRO, RoboCasa365 Target50, and RoboTwin C2R. Reported GPT-6 Astra results include 92.63% overall (741/800) across all eight LIBERO-PRO suites — versus 82.4% for RPent/Claude Opus 4.7 and 50.0% for the frozen π_RLinf VLA — and a chart-leading 59.20% on RoboCasa365 Target50; leaderboard published 2026-09-21. See also Case 25 for the underlying Harness VLA method.
• Astra on RoboMME, from Bingao Chen, Haoquan Fang, and C. Karen Liu (Stanford University): reports 79.13% success (633/800) across 16 tasks on the official RoboMME test set, averaging 3.63 Astra calls per episode. The three-tier controller uses a fine-tuned VLA and completion monitor; simulation is paused during planner inference. The blog includes methods, per-task results, and rollout videos; published 2026-09-22. See also Case 24 for the architecture.

• OpenAI: GPT-6 Astra: the original OpenAI announcement and system overview.

Credit belongs to the original authors, projects, and posts referenced by each case. This repository is a non-commercial compilation and technical analysis only; it does not claim ownership of the underlying work. Pull requests with new cases, corrections, and source updates are welcome.