Efficient Learning of Differentiable MPC Policies on the GPU via Smoothed ADMM and First-Order Reinforcement Learning
Efficient Learning of Differentiable MPC Policies on the GPU via Smoothed ADMM and First-Order Reinforcement Learning
This paper introduces SmoothedTurboMPC, a GPU-parallel differentiable model predictive control (MPC) solver that combines sequential quadratic programming with a log-barrier-smoothed alternating direction method of multipliers to address nonsmooth sensitivities and expensive evaluation. Paired with first-order reinforcement learning using Short-Horizon Actor-Critic, the approach reuses rollout MPC solutions to achieve up to 4.4x higher end-to-end training throughput. On a vehicle racing task, the method outperforms MPC PPO with a 46x speedup, training on 3.3 million transitions in 7 minutes. The framework is used to train a reliable MPC racing policy via RL with domain randomization and validated on a full-scale race car. The advances are integrated into the open-source TurboMPC solver.
The authors present SmoothedTurboMPC, a GPU-parallel differentiable MPC solver combined with first-order reinforcement learning (FORL) and Short-Horizon Actor-Critic (SHAC), claiming up to 4.4x higher training throughput and a 46x speedup on a vehicle racing task compared to MPC PPO, validated via domain randomization on a full-scale race car.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported duration
7 minutes
View original evidence
uses rollout MPC solutions instead of repeatedly re-evaluating them. Comparisons with state of the art solvers show that the approach achieves up to $4.4\times$ higher end-to-end training throughput. On a vehicle racing task, our method achieves higher returns than MPC PPO with a $46\times$ speedup, training on 3.3 million transitions in only 7 minutes. Finally, we use the framework to train a more reliable MPC racin
Open source S4speedup factor
46
View original evidence
uses rollout MPC solutions instead of repeatedly re-evaluating them. Comparisons with state of the art solvers show that the approach achieves up to $4.4\times$ higher end-to-end training throughput. On a vehicle racing task, our method achieves higher returns than MPC PPO with a $46\times$ speedup, training on 3.3 million transitions in only 7 minutes. Finally, we use the framework to train a more reliable MPC racin
Open source S4throughput multiplier
4.4
View original evidence
uses rollout MPC solutions instead of repeatedly re-evaluating them. Comparisons with state of the art solvers show that the approach achieves up to $4.4\times$ higher end-to-end training throughput. On a vehicle racing task, our method achieves higher returns than MPC PPO with a $46\times$ speedup, training on 3.3 million transitions in only 7 minutes. Finally, we use the framework to train a more reliable MPC racin
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
uses rollout MPC solutions instead of repeatedly re-evaluating them. Comparisons with state of the art solvers show that the approach achieves up to $4.4\times$ higher end-to-end training throughput. On a vehicle racing task, our method achieves higher returns than MPC PPO with a $46\times$ speedup, training on 3.3 million transitions in only 7 minutes. Finally, we use the framework to train a more reliable MPC racin
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org