[Guide] VLA models can already perform tasks, but real robots are still slow! PolicyTrim is a method to optimize VLA robot execution efficiency without retraining. It enhances task completion speed by extending reliable action sequences and eliminating redundant steps.
Vision-Language-Action (VLA) models unify visual observations, language instructions, and robotic actions into a single policy model, enabling robots to perform complex manipulation tasks based on natural language.
From OpenVLA and OpenVLA-OFT to π0.5 and GR00T, these models are emerging as key technological pathways in embodied intelligence.
But when the VLA model is actually deployed on a robot, a key bottleneck immediately becomes apparent: the total time required for the robot to complete a task depends not only on how fast each inference is, but also on how many inferences and how many physical actions the strategy requires.
Across multiple rollouts of the same task, the number of execution steps taken by the VLA model varies significantly, indicating that shorter, more direct successful paths are achievable—but the current policy rarely finds them.
Unreliable actions at the end of the chunk: The model predicts multiple actions at once, but later actions are more prone to accumulated errors, forcing the system to replan frequently. Forcing an extended execution length reduces success rates and increases the number of physical steps.
Excessive physical redundancy: Even if the task succeeds ultimately, the trajectory may include numerous corrective, backward, and repetitive motions.
Therefore, the deployment efficiency of VLA should not only be measured by the inference latency at each step, but also by policy efficiency: how many actions in a single prediction can be confidently executed? And how many actual physical steps are truly required to complete the task?
Most existing methods approach the problem from the computational side, such as visual token pruning, quantization, KV-cache reuse, distillation, or inference engine optimization. While these methods can reduce single-inference latency, if the strategy still requires numerous redundant physical steps, the actual task duration remains long.
Directly fixing longer action chunks is also unreliable.
The experiments in the paper show that as the action execution length is artificially extended, the success rate may decrease while the number of physical steps increases. This indicates that low-quality tail predictions introduce errors into the physical world, requiring the robot to perform additional corrective actions.
Key question: Can existing VLA policies be fine-tuned to automatically learn to "take fewer detours" and complete tasks with fewer physical steps, without compromising task success rates?
A team led by Professor Lei Yinjie from Sichuan University has proposed PolicyTrim—a two-stage reinforcement learning post-training framework focused on "policy efficiency." It does not alter the VLA architecture or collect new expert data, but instead further optimizes the robot’s real-world execution behavior based on existing pre-trained policies.
Project homepage: https://inceptionwang.github.io/PolicyTrim/
Paper link: https://arxiv.org/abs/2606.22540
Code link: https://github.com/INCEPTIONwang/PolicyTrim
Model link: https://huggingface.co/INCEPTIONwang/PolicyTrim
The new method enables:
- 3× action chunk utilization, more reliable execution over longer horizons;
- 51.4% reduction in physical steps, compressing redundant correction actions;
- 5.83× end-to-end maximum acceleration, LIBERO Object/π0.5;
From "Computing Faster" to "Doing Faster"
The core objective of PolicyTrim is not to replace computation-side acceleration, but to address another overlooked dimension: reducing the number of forward inferences required to complete a task. It decomposes policy efficiency into two optimizable goals: first, expanding reliable action chunks, then compressing redundant physical steps.
1. Reliable Action Chunk Extension
Many current VLA strategies output action sequences in the form of action chunks, but during deployment, teams often only dare to execute the first few steps. In its first phase, PolicyTrim employs dynamic execution horizon exploration: assigning different acceptance ratios to the same set of rollouts, enabling the model to concurrently explore reliable boundaries across short, medium, and long time windows.
Trajectories that successfully complete tasks with longer execution windows receive higher rewards, encouraging the model to make originally unreliable tail actions more usable.
The horizon reward is activated only when a successful trajectory occurs within the group, preventing the model from blindly pursuing longer chunk lengths in failure cases.
2. Redundancy-Aware Step Reduction
After the reliable horizon is extended, the second stage further reduces the total number of physical steps. PolicyTrim introduces a step-saving reward: trajectories that complete tasks with fewer steps receive higher rewards.
Incorporate "a shorter, more direct path to success" directly into the optimization goal.
Group-anchored regularization suppresses unreproducible speculative shortcuts, preventing the model from pursuing only short paths at the expense of success rate.
Two-stage sequential optimization prevents the two objectives—lengthening chunks and reducing steps—from interfering with each other, ultimately decreasing the number of forward inferences required to complete the task.
Experimental results
PolicyTrim was validated on LIBERO, ManiSkill, Meta-World, and real-robot tasks, covering various VLA architectures including π0.5, OpenVLA-OFT, and GR00T. Overall results show that PolicyTrim significantly improves execution efficiency while maintaining stable task success rates.
In LIBERO, π0.5 achieves up to 5.83x end-to-end acceleration while maintaining a success rate above 98%.
Across benchmarks, PolicyTrim achieves up to 2.36x speedup on ManiSkill and 2.52x speedup on Meta-World.
Across architectures, the retrained OpenVLA-OFT achieves a 2.97x speedup using the full two-stage pipeline; OpenVLA using only the second stage achieves a 1.41x speedup.
Qualitative comparison: Fewer detours, a more direct path
In randomly sampled LIBERO tasks, the baseline often exhibits redundant correction actions and hesitation/jittering near the target; PolicyTrim produces smoother, more direct trajectories and completes the same tasks in fewer physical steps, typically reducing the number of physical steps by about half.
Real robot deployment: Efficiency gains from simulation can be transferred to the physical world.
The paper further validates real robotic tasks on an Agilex Piper manipulator equipped with two Intel RealSense D435i cameras, including three types of tasks: FlipMug, HangMug, and TapeBox. Experiments cover both a standard setup with fixed target positions and a dynamic setup with random perturbations to the target during grasping.
PolicyTrim maintains or improves success rates under both standard and dynamic perturbation settings while significantly reducing real-world execution time. Under standard real-robot settings, it achieves an average end-to-end acceleration of 1.86x.
The experimental results show that PolicyTrim does not merely optimize benchmark metrics: on physical robots, task completion time is reduced to approximately half that of the baseline, while maintaining robustness in dynamic interference scenarios.
Why does PolicyTrim work?
• Efficiency improvements for the strategy itself: It reduces redundant actions and unnecessary replanning, not just single inference latency.
• No need for training from scratch: As a post-training framework, it can further optimize existing VLA models, reducing data collection and training costs.
• Balance speed and success rate: Reliable horizon expansion avoids blindly extending chunks, while stability constraints prevent short-path speculation.
• Can be combined with computational acceleration: PolicyTrim optimizes the number of forward inferences, while methods like VLA-Cache optimize the time per inference; these two approaches are orthogonal and can produce compounded acceleration.
The core insight of PolicyTrim is that for VLA robots, deployment efficiency comes not only from faster model inference but also from more efficient policy behavior—helping the robot "take fewer detours" can lead to significant speed improvements.
Reference: https://arxiv.org/abs/2606.22540
This article is from the WeChat public account "New Intelligence Yuan," authored by New Intelligence Yuan; edited by LRST.