AI

RoboSPA: Pick the Third-Farthest Bottle, Then Remember What You Hid

Shar Hendrix 4 min read

“Pick the ketchup” is easy. “Pick the third-farthest pill bottle from the brown one” is the test. RoboSPA (Robot Spatial-Procedural Assessment), arXiv:2609.05324, landed 4 September and is accepted at EMNLP 2026. Zhejiang University and partners built it on RoboTwin 2.0.

The corpus: 527K trajectories, 997 hours of interaction video, 108 million timesteps, 56 base tasks, 280 variants, 5 embodiments (Aloha-AgileX, ARX-X5, Piper, Franka, UR5). 463K of those runs are domain-randomized.

RoboSPA overview of ten task categories with dual-arm table scenes and language prompts
Ten capability buckets, from geometric attributes to memory. Source: RoboSPA project page.

Two axes, five notches

Fine-grained spatial reasoning covers geometric attributes, distance, row-column indexing, referential relations, and cross-view prompts. Long-horizon procedural planning covers repeats, order-free piles, ordered sequences, composite action chains, and memory (hide the blocks, then match the ones in back).

Each base task has levels L1–L5. Spatial variants add lookalike candidates. Procedural variants add steps. Metrics besides success rate: ONTA (chance-corrected target pick) and Progress Score (fraction of subgoals done).

Paper protocol on clean Aloha-AgileX: 50 successful expert demos per variant, train on all five levels (250 demos) per base task, 100 eval rollouts with unseen instructions.

RoboSPA taxonomy wheel plus rows of increasing object count and longer action sequences
Difficulty is extra objects or a longer sequence, not a new task name. Source: RoboSPA project page.

L5 is a cliff

Single-task results on clean scenes:

ModelL1 successL5 success
RDT16.8%6.9%
GO-125.1%8.8%
π0.555.2%22.3%
X-VLA50.4%19.9%

Every evaluated model stays under 25% overall at L5. Spatial ONTA sits near chance when the cue is a relation, not a category. Progress Score beats final success: the policy starts the recipe and dies on a later step. Memory-intensive L5 is a wipe.

Failure examples for grounding, manipulation, ordering, repetition, and memory on dual-arm table tasks
How the models fail, not only whether. Source: RoboSPA project page.

Code, configs, and a Hugging Face dump are up. This is a diagnostic, not a leaderboard for warehouse humanoids. The arms are tabletop dual-arm setups.

A Human’s Take

I am tired of 90% on pick-the-red-cup. Asking for the fourth object in row one, counted right to left, is closer to how people actually talk to a robot. If π0.5 is the high-water mark at 22% on L5, the model is still guessing which bottle you meant.

Sources