Users are excited about ELASTIC's RL-optimized diffusion robot policies because the meta-policy execution appears magical in fluidly trading off sequential and parallel compute, and they praise the collaborator's thoughtful contributions.
Based on 2 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Equally magical was @andrewzouli, who impressed me at every stage of this project, from his depth of thought, clarity of communication, and finesse of execution -- it was much too easy to forget he isn't a PhD student yet! Anyhow, check out https://arxiv.org/abs/2606.31132 for more! [6/6]
Watching the meta-policy trained via ELASTIC execute is sometimes magical -- we see it fluidly trade off sequential and parallel compute depending on how close to the object the gripper is / the multi-modality of the base policy. [5/n] https://x.com/g_k_swamy/status/2076744499934667107/photo/1
Very excited about this new paper:). It's well known that test-time compute (more sequential denoising steps or parallel samples) can improve policy performance. However, knowing *when* and along *which axis* to scale is often unclear, requiring interacting with the policy / world. We introduce ELASTIC: a principled and general RL algorithm for learning optimal test-time compute allocations for generative control policies!
While more compute is usually better, the real-world latency requirements of robots mean we need to think carefully about *where* and *how* we scale. What makes this hard is that it is often unclear a priori what kind of scaling (if any) will improve performance. [2/n]
Equally magical was @andrewzouli, who impressed me at every stage of this project, from his depth of thought, clarity of communication, and finesse of execution -- it was much too easy to forget he isn't a PhD student yet! Anyhow, check out https://arxiv.org/abs/2606.31132 for more! [6/6]
Watching the meta-policy trained via ELASTIC execute is sometimes magical -- we see it fluidly trade off sequential and parallel compute depending on how close to the object the gripper is / the multi-modality of the base policy. [5/n] https://x.com/g_k_swamy/status/2076744499934667107/photo/1
Very excited about this new paper:). It's well known that test-time compute (more sequential denoising steps or parallel samples) can improve policy performance. However, knowing *when* and along *which axis* to scale is often unclear, requiring interacting with the policy / world. We introduce ELASTIC: a principled and general RL algorithm for learning optimal test-time compute allocations for generative control policies!
While more compute is usually better, the real-world latency requirements of robots mean we need to think carefully about *where* and *how* we scale. What makes this hard is that it is often unclear a priori what kind of scaling (if any) will improve performance. [2/n]
Users are excited about ELASTIC's RL-optimized diffusion robot policies because the meta-policy execution appears magical in fluidly trading off sequential and parallel compute, and they praise the collaborator's thoughtful contributions.
Based on 2 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.