AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Differentiable Model Predictive Control on the GPU

Emre Adabag, Marcus Greiff, John Subosits, Thomas Jonathan Lew

ICLR 2026 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies · 9 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace Riccati recursion with parallel PCG solver (GPU)DiffMPC vs trajax · RL timing, Problem 1 (GPU)backward-pass time (s): 1.83 -> 0.32Improved
DiffMPC vs trajax on nonlinear attitude stabilizationtrajax · Nonlinear attitude stabilizationbackward time (ms): 505.0 -> 69.3Improved
run DiffMPC on CPU instead of GPU (vs trajax)DiffMPC · RL timing, Problem 4 (CPU)backward-pass time (s): 5.95 -> 10.00Hurt
warm-start PCG with previous solution (ablation)DiffMPC · RL timing, PCG tol 1e-12forward speedup: 4%Improved
warm-start PCG with looser tolerance 1e-4 (ablation)DiffMPC · RL timingforward speedup: 4% -> 11%Improved
use DiffMPC instead of trajax for imitation learningtrajax · Cart-pole imitation learningtraining time: approximately 2 times speedupImproved
RL-tune MPC cost and friction parameters with domain randomizationBaseline MPC · Drift with water puddles (simulation)success rate (%): 70% -> 100%Improved
modify mpc.pytorch to remove serial list comprehensionmpc.pytorch · RL timing, Problem 1 (GPU)forward time (s): 67.9 -> 1.9Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy