Robot dev claims GPT6 as policy stacks boxes zero-shot at near-100% success
builds_robots · x · 2026-09-21
Robot developer buildsrobots demoed an aggressive approach: using GPT6 directly as a robot policy that outputs end-effector-frame actions from two RGB-D cameras and instructions, with no task-specific training.
Two test tasks:
- "Find a wooden box (can be obscured) and put it on the yellow sign"
- "Stack the wooden boxes on each other"
The author claims amazing performance — near 100% success given unlimited tokens and retries. In follow-up discussion, ChongZzZhang noted the model can deliver individual skills and chain them via prompting, though a single task may take 30 minutes; token costs are very high and more dexterous tasks like folding boxes remain doubtful.
Caveat: this demo is from an individual developer, unverified, and the existence of "GPT6" cannot be confirmed.
Related event: GPT6 as robot policy: zero-shot bin stacking with wrist cameras only(5 posts)→
More from Embodied
- Waymo service has gotten significantly worse in recent months, reports regular rider — iScienceLuvr · 2026-11-12
- AI smart ring promises memory, app control and health tracking, preorders open — rohanpaul_ai · 2026-11-12
- Scoble: X is underestimating Meta, Google, Snap and Chinese AR glasses — Scobleizer · 2026-11-12
- Otto Pro: A Plug-and-Play Desktop Box That Runs AI Agents Locally, Keeping Your Data at Home — heyshrutimishra · 2026-11-12
- NVIDIA's Isaac Orbit Becomes Isaac Lab, Powering GR00T Robot Training Pipelines — animesh_garg · 2026-11-12
- Abbeel lab's SuperMimic replaces physics sim with differentiable world model, ~1000x faster sim-to-real — HirokatuKataoka · 2026-11-12