thijs — Astra paints the Golden Gate Bridge with a real robot arm

TL;DR

thijs (@cdngdev) hands OpenAI’s GPT-6 Astra a robot arm, a paint brush, and a camera and asks it to paint the Golden Gate Bridge in real life. Astra figures out how to control the robot, plans one minute of actions at a time and monitors execution in the background to make in-flight adjustments, and progressively improves across attempts — the sharer feeds occasional “what can you do better?” prompts and Astra updates its own approach. Framed by the author as “algorithmic art, similar to how pen plotter is.” A five-day-after-launch data point for GPT-6 Astra as a direct real-hardware robot controller, with an explicit self-critique loop closing the gap between attempts rather than a trained policy.

Key claims

Astra controls a real robot arm holding a paint brush from a camera feed, with no fine-tuned policy or teleop demonstrations — the LLM is the controller [Post 1/thread, Post 3/thread].
The control loop is structured: Astra plans one minute of actions at a time and monitors execution in the background to make adjustments, both between and sometimes during runs [Post 3/thread].
Between runs the human occasionally supplies natural-language feedback (“what can you do better?”) and Astra updates its own approach — an in-the-loop RL-of-language pattern with the human closing the credit-assignment step [Post 3/thread].
Progressive improvement across attempts is visible in the timelapse — later attempts are qualitatively better paintings than earlier ones [Post 4/thread].
Author frames the result as algorithmic art — comparable to pen-plotter art rather than a replacement for human artists [Post 2/thread].

Method

Tweet thread with a 68-second timelapse video (Post 1), a philosophical follow-up on algorithmic art vs. human artists (Post 2), a description of the control loop and human feedback pattern with two supporting stills (Post 3), and a “progression is incredible” side-by-side video (Post 4). No paper, no code, no numbers. The robot is a small tabletop arm holding a paint brush; the target is a physical painting of the Golden Gate Bridge on an easel. Astra is OpenAI’s GPT-6 model, launched Sep 3, 2026 (five days before this thread). No comparison against Claude / Gemini / prior GPT models on the same rig, no report of tokens, cost, per-attempt latency, or how many attempts the timelapse covers.

Results

Qualitative only. One completed painting of the Golden Gate Bridge produced by iterated attempts under a natural-language “what can you do better?” feedback loop; the sharer’s characterization is that Astra “progressively got better throughout its attempts.” No metric captured — no similarity score against a reference, no per-attempt success/failure judgment, no independent evaluator, no wall-clock or token budget. Third-party reply photo (from @harinsrikanth) appears to show the painting alongside the physical bridge.

Why it’s interesting

Sits directly in the LLM-as-robot-controller cell of the recipe-lever board on VLA Models and the physical-tool-use cell on Tool-Use Agents — one week’s worth of independent hobbyist reproductions of the pattern that Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK measured formally on YAM arms (Astra ties Fable 5.1 at 3.9× fewer tokens) and that Sentdex — general multimodal LLMs may sidestep VLA/WAM research for high-level robotics intelligence showed with open-weights GLM-5.3-Flash on an XGO quadruped. Two moves make this thread worth filing despite the absence of numbers: (i) the plan-one-minute-then-monitor control structure the author describes is a concrete implementation of the async-planning + interruptible-tool pattern VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation proposes for embodied orchestration, at frontier-LLM scale with no VLA underneath — the robot arm is being driven end-to-end by an LLM emitting plans against a camera feed rather than an orchestrator over a policy; (ii) the human-in-the-loop natural-language critique between attempts (“what can you do better?” → Astra updates its own approach) is a hand-executed version of the learned-experiences-reduce-required-interaction-depth finding from Training-Free GRPO (Training-Free Group Relative Policy Optimization) applied to a physical-world task, and prefigures what Anthropic’s Model Hardware Standard (Previewing the Model Hardware Standard) is building infrastructure for. Note the caveats consistent with the rest of the LLM-as-controller thread: no baseline, no evaluation protocol, small-sample demo by a launch-day tester; and the artistic framing (“algorithmic art”) sidesteps the harder question of whether the same setup handles contact-rich functional manipulation, which is exactly where prior Robocurve runs (Gemini Robotics-ER 2 vs Claude Opus 5 clapboard eval — ER2 threw robot arms off table, hallucinated success (Robocurve / Jay Chooi), Gemini 3.7 Flash saturates Robocurve physical tool-use benchmark at 92% vs 32% for 3.6 Flash (three weeks apart)) show LLM controllers struggle most. Complements the Thomas Ricouard — Astra builds a Blender-to-UE5 walkable house end-to-end via Python 3D-as-code Astra-drives-Blender-and-UE5 thread as the physical counterpart to that same-day virtual code-agent demo — same model, same launch window, one working through Python DCC tools, the other through a real end effector.