yup, suprisingly audio codec models are good action tokenizer recipes!
will presenting this work at corl in november!
Two of our arms striking a match, fully autonomous. Ahad Jawaid pulled this off for his CoRL 2026 paper on our previous generation arms.
Our 2026 lineup is built for even more dexterous work, with academic discounts for research labs: buff.ly/nNrnn7K
i’m going to be presenting this work at CoRL in Austin on novemeber 9-12th!
would be cool to chat with people who are working on scaling robot foundation models and multimodal tokenization
Starmind
spacex.com/spacexai/starmind…
introducing hierarchical implicit q-chunking (HIQC)
long-horizon offline reinforcement learning is hard
the more steps an agent takes, the more value estimation errors compound
i reduce this by decomposing the horizon into smaller segments with subgoals and action chunks
this simple idea leads to an increase of nearly 10%+ average compared on 9 long horizon simulation task xitter.cf/ahadj0/status/20723344…
i found by simply using a high level policy predicting latent subgoal and a low level policy prediciting action chunks
you can reduce the number of bootstrapped value estimates to O((T/k)^(1/2)) xitter.cf/ahadj0/status/20723344…
i’d like to thank @seohong_park for open sourcing his work as this work was heavily based on those prior works
project: ahadjawaid.com/hiqc
paper: ahadjawaid.com/hiqc/59_Offli…
ahad retweeted
Check out Ahad's work on leveraging neural audio codec architectures for VLA action tokenization.
An interesting direction that adapts ideas from audio compression to represent continuous robot actions.👇
i will like to thank the authors of the OAT paper as our codebase was developed on top of theirs @liu730chaoqi and my co author @YuXiang_IRVL for making this work possible
paper: arxiv.org/abs/2606.21372
code: github.com/ahadjawaid/nac
project page: ahadjawaid.com/nac
also want to acknowledge @jeluna20 as a core contributor on this project, and the open source repos it’s built on: @lerobothf @uma_robots @vectorwang2







