HKUST(GZ) & OLA

SLA-Preview↗

SLA-Preview is a VLA from HKUST(GZ) & OLA, released November 5, 2026. It ranks #2 of 36 overall on the Robot Intelligence Index, with a rating of 1807. Its best result is #1 of 14 on RoboDojo Real. A spatial-language-action model: π0.5 continually pre-trained on about 1.7M robot episodes with alignment to a frozen 3D perception model. Its RoboDojo real-robot checkpoint adds nine past camera frames as memory.

Release date: Nov 05, 2026

Type
VLA
Model weights
Open
License
Custom license
Parameters
—
Pricing
Self-hosted
Benchmarks
2
Overall rank
#2of 36
Rating 1807 · range 1–8
Real robots rank
#4of 56
Rating 1837
Simulation rank
Not ranked
Needs two simulation benchmarks; has one (RoboTwin 2.0).
Rating
1807± 10.1
10601989
Cost / task · est.
$0.041
$0.0087$17.74
Time / task · est.
45s
9.5 s1 h 22 min
Rank on each benchmark
BenchmarkScoreRank
RoboDojo RealOverall42.31 / 14RoboTwin 2.0Overall79.28 / 36

Per task attempt; the tiles show the median across benchmarks. Self-hosted compute is priced at one H100 at $3.29 per hour (Lambda on-demand, Oct 2026).

Inference

Not reported. Spatial alignment modules are training-only, so inference architecture is identical to the pi0.5 policy it is initialized from.