GPT6 as Policy: Wrist-Mounted Dual Cameras Drive Zero-Shot Robot Stacking
On September 21, robotics developers ChongZzZhang and buildsrobots released a demo using "GPT6 directly as a robot policy": the model takes feeds from two wrist-mounted RGB-D cameras plus natural-language instructions and end-to-end outputs end-effector (ee-frame) actions, with no specialized training. Tasks included bin-stacking operations like "find the wooden crate (which may be occluded) and place it on the yellow marker," with buildsrobots claiming a success rate near 100%.
Confirmed
- The system relies on no base/world-frame observations and no gripper feedback, completing spatial manipulation purely via closed-loop vision
- Camera mounting and action format were described only by a Claude-generated prompt, with no dedicated harness used
- The demo video is played at 4x/12x speed
Why it matters
- ChongZzZhang distilled two takeaways: closed-loop spatial understanding is now an off-the-shelf capability; reasoning and affordance can be achieved without symbolic representations
- If "GPT6 as policy" can truly drive real-world robot manipulation zero-shot, it challenges the traditional view that robot policies require dedicated training
- Successfully grasping and placing occluded targets suggests the model has some closed-loop error correction, rather than open-loop action replay
- The "near 100% success rate" is a one-sided claim by buildsrobots; sample size and test conditions were not disclosed
2026-09-21 ~ 2026-09-21 · 5 related posts
Primary sources
- Robot dev claims GPT6 as policy stacks boxes zero-shot at near-100% success — builds_robots ·
- Closed-loop spatial understanding from just two wrist cameras, no gripper feedback — ChongZzZhang ·
- No-harness stereo demo: prompt-only camera setup shows closed-loop spatial understanding is off-the-shelf — ChongZzZhang ·
- GPT6 as robot policy: end-to-end ee-frame actions from dual RGB-D, near-perfect box stacking — ChongZzZhang · 2026-09-21
- Pure-vision robot policy: two wrist RGBD cams + claimed "GPT6" hits near-100% success — ChongZzZhang · 2026-09-21
- [source] Closed-loop spatial understanding from just two wrist cameras, no gripper feedback — ChongZzZhang · 2026-09-21
- [source] Robot dev claims GPT6 as policy stacks boxes zero-shot at near-100% success — builds_robots · 2026-09-21
- [source] No-harness stereo demo: prompt-only camera setup shows closed-loop spatial understanding is off-the-shelf — ChongZzZhang · 2026-09-21