PathTime-VLA: Path-Time Decoupling for Factorized Post-Training of Vision-Language-Action Policies

arXiv:2610.11771 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "PathTime-VLA: Path-Time Decoupling for Factorized Post-Training of Vision-Language-Action Policies".

Rosa: The gist The PathTime-VLA framework introduces a path-time decoupled action representation for Vision-Language-Action policies, separating geometric guidance from execution timing to enable distinct post-training interfaces for geometry and timing.

Dev: First, who's behind it and why it matters.

Paper summary: Rosa: So, to recap what we just heard about "PathTime-VLA: Path-Time Decoupling for Factorized Post-Training of Vision-Language-Action Policies," the paper’s main thesis is about separating how a robot moves geometrically from how fast it moves through that path.

Dev: They claim that current vision language action models usually predict actions at fixed time intervals, which means the route and the pace are coupled together, making adaptation really complicated when you have things like interface delays or human operator behavior involved.

Taro: That coupling is what causes issues because useful geometric guidance often comes with a timing that’s shaped by those delays, and they found a way to bring path-time parameterization from classical motion planning into the learned action representation.

Rosa: They introduce PathTime-VLA, which uses a progress-indexed interaction path and an interval-time profile to define a monotone time law t(s), which produces controller commands at specific times along that path, like X(s(t)) <ref:2610.11771#pg6>.

Dev: The key insight is that for one given geometric path, you can express different executions just by changing that time profile, letting you choose chunk-wise speed adjustments without changing what the final geometric prediction target is.

Taro: This representation lets them have alternative executions expressed through the time profile, which allows for chunk-wise speed choices while keeping the geometric prediction target exactly the same.

Rosa: This structure supports a staged post-training procedure where demonstrations and DAgger interventions first establish and expand a supervised prior for target domain manipulation.

Dev: Then Speed-DQN learns adaptive chunk-level pacing while keeping the original path generator fixed, which is an interesting constraint to put on the learning process.

Taro: Finally, Path-AWR refines the path generation from rollout outcomes while still retaining that learned timing policy, so they’re separating geometry and timing learning interfaces.

Rosa: This means geometry and timing are post-trained through different learning signals suited to their specific roles, but the overall behavior is still judged under the same task-level outcomes.

Dev: It’s about making sure you don't confuse the geometry training signal with the execution speed training signal, which I think is a smart way to approach it.

Taro: If you only listen to this show, what it means is that complex robots are moving better because they can independently optimize their path following and their actual speed control during adaptation.

Rosa: So they’re showing how this decoupling helps the system handle the messy reality of execution timing without sacrificing the geometric accuracy of the plan itself.

Conclusion: Dev: So we've covered the core idea of this PathTime-VLA paper, which is about that path-time decoupling for factorized post-training policies Qing Huang, Yifei Yang, Ziqing Zou, Anzhe Chen, Zhenjie Zhu, Yufei Wei and Rong Xiong.

Rosa: The title points to the main contribution: separating the progress-indexed interaction path from its interval-time profile to create distinct interfaces for geometry and timing.

Taro: What this means in simple terms is that instead of having one single way to train a robot policy, you have two specialized ways for learning—one focused on where it goes, and one focused on how fast it gets there.

Dev: Think about the implications: this structure lets them use different types of learning signals for those two things during adaptation.

Rosa: It means that geometry gets feedback tailored to spatial accuracy, while timing gets feedback based on real-world execution performance metrics from interacting with the robot.

Taro: For someone listening who isn't a robotics expert, this means the system is much better at adapting its pace when things go wrong during deployment because it can learn the speed correction without messing up the overall plan geometry.

Dev: The authors are showing that this staged post-training procedure, involving DAgger and Speed-DQN, actually works to refine both components in a coordinated way.

Rosa: And the results on three manipulation tasks show that this method achieves a high success rate and a lower mean completion time over successful trials compared to other tested variants.

Taro: So ultimately, it supports path-time factorization as a practical interface for reliable adaptation and efficient execution in complex real-world robot applications.

Zhejiang University

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: 8 pages, 7 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The gist The PathTime-VLA framework introduces a path-time decoupled action representation for Vision-Language-Action policies, separating geometric guidance from execution timing to enable distinct

Key concepts

PathTime Representation
Motion is represented by an interaction path X(s) and a positive interval-time profile t(s). This structure keeps the geometric progress fixed while allowing different execution timings. It lets the system choose chunk speeds without changing the target geometry for manipulation.
Factorized Post-Training
The method creates separate learning signals for geometry and timing. Demonstrations and interventions provide corrective targets, while task outcomes evaluate the resulting executions. This allows different stages—supervised adaptation, RL exploration, and diffusion refinement—to learn complementary aspects independently.
SpeedDQN
This component learns adaptive chunk-level pacing by selecting execution multipliers based on task feedback. It uses a reward that balances terminal task success with a preference for faster bins. This helps the policy choose the right speed for different parts of the motion path.
Path-AWR
Path-AWR uses evidence from SpeedDQN to refine the diffusion path generator. It trains a critic and baseline on execution outcomes, where weights depend on ordering rather than just advantage scale. This transfers outcome information back to improve the geometric path generation.

Terminology

Summary

The gist The PathTime-VLA framework introduces a path-time decoupled action representation for Vision-Language-Action policies, separating geometric guidance from execution timing to enable distinct post-training interfaces for geometry and timing.

How it works

  1. PathTime-VLA represents motion as a progress-indexed interaction path X(s) and a positive interval-time profile, which defines a monotone time law t(s), yielding controller commands X(s(t)) <ref:2610.11771#pg6> The representation separates fixed-progress path samples from interval durations defining t(s) For a given path, alternative executions are expressed through the time profile, allowing chunk-wise speed choices without changing the geometric prediction target This representation supports a staged post-training procedure: demonstrations and DAgger interventions establish a target-domain prior, SpeedDQN learns execution multipliers from robot interaction, and Path-AWR uses rollout outcomes to refine the diffusion path generator A path-conditioned action expert realizes the resulting motions while maintaining distinct learning interfaces for path generation and execution timing.

Path-Time Action Representation

The method constructs a chunk by path samples and their traversal times, where Xk = xk(σk,i) for i=1 to K The discrete gripper command does not contribute to the SE(3) progress metric, as at each path knot, it retains the latest command at its corresponding recorded time and uses zeroorder hold during composition. The TimeHead uses cross-attention to distribute traversal time across path intervals from the context, predicted path, and target-rate coordinate A scalar multiplier alone would uniformly compress or expand a chunk, which cannot express the phase-dependent timing needed for contact-rich manipulation.

Factorized Post-Training

The representation provides separate interfaces for complementary learning signals: demonstrations and interventions supply dense corrective targets, while task outcomes evaluate the resulting executions The three stages in Fig. 3 use supervised adaptation to establish a target-domain prior, low-dimensional RL to explore timing choices, and advantage-weighted diffusion training to transfer outcome information back to path generation.

Speed-DQN and Path-AWR

SpeedDQN learns adaptive chunk-level pacing by selecting execution multipliers mk = m(bk) from the ordered task-specific set Mτ = m(b) for b=0 to B−1. The reward combines terminal task feedback with a bounded preference for faster bins. Path-AWR uses this evidence to refine the diffusion generator by training a path critic and its state-value baseline on Rexec, where weights depend on ordering, not advantage scale.

Experimental Results

Across three manipulation tasks—Peg insertion, Microwave opening, and Tape placement and drawer closing—the complete method achieves 58/60 successes versus 57/60 for PathTime-VLA under BC + DAgger at fixed 1×. The mean completion time over successful trials is lower for PathTime-VLA in all six task–supervision comparisons, supporting the representation before SpeedDQN or Path-AWR. The complete method achieves the highest throughput proxy among the evaluated fixed-rate and reinforcement variants on each task under BC + DAgger.

Conclusion

The representation keeps the geometric path target invariant to timing changes while exposing execution timing as a compact control space. This supports a clear division of learning signals: DAgger establishes a target-domain supervised prior, Speed-DQN improves timing from task feedback, and Path-AWR transfers outcome information back to the PathHead. Across 600 real-robot evaluation trials on three manipulation tasks, the complete method achieves 58/60 successes and the highest throughput proxy on each task among the evaluated fixed-rate and reinforcement post-training variants. Compared with PathTime-VLA under BC + DAgger at fixed 1×, it achieves a similar observed aggregate success rate and reduces mean completion time over successful trials by approximately 39–52%. Together, they support path-time factorization as a practical interface for reliable adaptation and efficient execution >.

Improvements for AI systems

  1. Bold Header: Path-Time Decoupled Action Representation

The system can now separate a progress-indexed interaction path from its interval-time profile, which provides distinct post-training interfaces for geometry and timing. This allows for independent optimization of spatial guidance (geometry) and execution pace (timing).

  1. Bold Header: Speed-DQN for Adaptive Chunk Pacing

The AI can now learn adaptive chunk-level pacing by using low-dimensional online reinforcement learning with a fixed action expert. This enables the robot to dynamically adjust its speed based on task outcomes, as seen in the reward function that has a bounded preference for faster bins.

  1. Bold Header: Path-AWR for Outcome-Guided Path Refinement

The system can now refine path generation from rollout outcomes using an advantage-weighted learning approach. This means the diffusion model is updated based on executed paths, specifically using advantage-weighted regression, to ensure refined paths outperform the context-dependent baseline.

  1. Bold Header: Enhanced Adaptation via Factorized Post-Training

The complete method supports a staged post-training procedure where demonstrations and DAgger establish a target-domain prior, followed by speed optimization, and finally path refinement, resulting in a policy that has exposes distinct post-training interfaces for geometry and timing.

  1. Bold Header: Improved Execution Through Precision Timing

The robot can exhibit phase-dependent motion control, as the TimeHead can move rapidly in free space, decelerate before contact, dwell during engagement, and accelerate after release while meeting the requested global pace. This overcomes limitations of fixed-interval policies by allowing for contact-rich manipulation timing.

Sources

Related papers