GPT6 as robot policy: zero-shot bin stacking with wrist cameras only
On Sept 21, roboticists ChongZzZhang and buildsrobots demoed using GPT6 directly as a robot policy: the model end-to-end outputs end-effector-frame actions from two wrist-mounted RGB-D camera streams and natural language instructions, with no dedicated training, completing bin-stacking tasks; buildsrobots claims near-100% success.
Confirmed
- Tasks include finding a wooden bin (possibly occluded) and placing it on a yellow marker; success on occluded targets suggests closed-loop error correction rather than replayed trajectories
- The system uses no base/world-frame observation and no gripper feedback, relying purely on visual closed-loop control
- Camera mounting and action format were described only via a Claude-generated prompt, with no specialized harness
- Demo video plays at 4x/12x speed
Unconfirmed
- The near-100% success rate is a single-party claim by buildsrobots; sample size and test conditions undisclosed
Why it matters
- ChongZzZhang draws two conclusions: closed-loop spatial understanding is an off-the-shelf capability, and reasoning plus actionability can work without symbolic representations
- If verified, zero-shot general multimodal models could replace specially trained robot policies
2026-09-21 ~ 2026-09-21 · 5 related posts
Primary sources
- [source] 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
- 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
- No-harness stereo demo: prompt-only camera setup shows closed-loop spatial understanding is off-the-shelf — ChongZzZhang · 2026-09-21