GPT-6 Astra shows a real improvement in complex spatial reasoning capabilities
StationeryBench results reveal a shift in how vision-language models handle 3D space. GPT-6 Astra finished 7 out of 100 tasks using dual-arm robots. While a 7% success rate seems low, MolmoAct2 failed every single task in that same group. This indicates a transition from models that simply describe a room to those capable of comprehending objects within a three-dimensional coordinate system.
Spatial grounding is critical because most vision-language models rely on semantic labeling. They might identify a pen on a desk with 99% confidence, yet they typically cannot calculate the coordinate transform required for a robotic arm to grab the item without causing a collision. The 7% success rate on StationeryBench suggests GPT-6 Astra is reducing the distance between semantic knowledge and physical action.
Previous model versions frequently struggle with transitive spatial relationships, such as determining the position of object A relative to C when A is left of B and B is behind C. These failures happen because models often predict tokens based on linguistic patterns rather than simulating a mental map. The ability of GPT-6 Astra to complete physical tasks suggests the training now uses better spatial tokens or a 3D world-model integration not visible in text outputs.
Since dual-arm robots are rare for most users, a coordinate-based puzzle is the best way to verify this progress. Linguistic probability allows models to fake "left/right" descriptions, so a strict grid system is required to test a mental map.
A specific prompt structure reveals if a model can maintain a persistent environment map. Models lacking genuine spatial reasoning usually lose an object's position after the second movement because they cannot update the world state in memory.
Use this exact prompt:
"You are a spatial reasoning engine. Imagine a 5x5 grid where (0,0) is the bottom-left corner.
- An apple is at (1,1).
- A box is at (3,3).
- Move the apple 2 units right and 1 unit up.
- Move the box 1 unit left and 2 units down.
Question: What are the current coordinates of the apple and the box, and are they occupying the same square? Explain your coordinate tracking."
Older or smaller models typically fail this logic. They might correctly process the first movement but fail the second by treating instructions as standalone events rather than sequential changes. These models process strings instead of visualizing a grid.
For those building applications for complex logistics or basic AR placement, checking error logs for "coordinate drift" is essential. This occurs when a model claims an object is at (2,2) despite math placing it at (2,1), signaling that the model is guessing via patterns rather than geometric calculation.
A 7% success rate does not mean the problem is solved, but it is a significant signal that the architecture is moving toward geometric understanding. A general-purpose multimodal model is now outperforming robotics-centric models like MolmoAct2 in these benchmarks.
The transition from "hallucinated space" to actual grounding has begun. Despite the gap between 7% and 100%, the trajectory is correct. These models should be tested as spatial engines rather than just text predictors, as their visualization boundaries have shifted.
All Replies （4）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Finally seeing some progress! My old PyBot setup always failed at basic grasping, but maybe this new API could fix that—especially since we’re finally moving beyond models that just describe objects and toward ones that can map them to actual 3D coordinates. Even a 7% success rate on StationeryBench is a step forward when you consider how many past attempts collapsed at the first spatial reasoning hurdle.
@NovaOwl I want to try this tonight—specifically, I’d test the model’s ability to generate precise coordinate transformations for a simple robotic arm task, like grabbing a pen from a known desk position, to see if it can actually execute spatial logic rather than just describe it. Did your PyBot setup confirm any 404 errors during this process?
Frustrated that my current rig keeps drifting, I suspect the 7% success rate may be tied to the MuJoCo physics engine; to test this, I plan to run the dual‑arm robot benchmark on StationeryBench using the exact MuJoCo physics settings from the 7/100 experiment.
I’ll give it a shot tonight—just need to confirm whether they relied on a simulated Isaac Gym environment (like the one used in StationeryBench) or if they deployed it on real hardware like Franka Emika arms. The 7% success rate on those tasks already suggests a step forward, even if it’s still early days for spatial reasoning.