MobileVISTA is introduced as a follow‑up to VISTA. It augments demos in the image space to improve robustness to robot base position changes. The claim is that policies trained from one base pose can work from many poses despite base pose errors of a few centimeters.
Key points
MobileVISTA follows previous work VISTA.
It aims to collect demos at one base pose and obtain a policy that works from many poses.
Sources
SLAM stack consistency over ~400 m, GPS optional
The SLAM stack stays consistent from the 3rd floor, down a staircase, and out onto the street over ~ 400 metres. GPS is optional now. The author reports barely any observable drift.
Key points
GPS optional now
Consistent SLAM over ~ 400 metres with barely any observable drift
Sources
New GPU DSL built on OxCaml matches cuBLAS on H100
A new GPU domain-specific language is being started, using OxCaml’s high-performance programming model. Version 0 of its compiler can express an 8192³ TF32 GEMM in 100 lines of OCaml and matches cuBLAS performance on an H100.
Key points
DSL: GPU DSL built on OxCaml’s high-performance programming model.
Benchmark: 8192³ TF32 GEMM in 100 OCaml lines matches cuBLAS on an H100.