PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Qing Huang, Yifei Yang, Ziqing Zou, Anzhe Chen, Zhenjie Zhu, Yufei Wei, Rong Xiong, Yue Wang

arXiv:2610.11771v1cs.RO

TL;DR

Fixed-time VLA actions couple a robot's route with its execution pace, while teleoperated timing can reflect interface delays and operator behavior. PathTime-VLA separates a progress-indexed path from its timing profile and post-trains each with distinct signals. Across three tasks, the complete method achieved 58/60 successes and approximately 39–52% shorter mean completion times than its BC + DAgger fixed-1× comparison.

  • Problem

    Teleoperated demonstrations provide useful spatial guidance, but their timing may reflect interface delays and operator behavior rather than an efficient autonomous schedule.

    Source: §I · p. 1

  • Method

    PathTime-VLA separates a progress-indexed path from its interval-time profile, using demonstrations and DAgger for a target-domain prior, Speed-DQN for timing, and Path-AWR for rollout-guided path refinement.

    Source: §I · p. 2

  • Results

    Across 600 real-robot trials, the complete method achieved 58/60 successes and approximately 39–52% shorter mean completion time over successful trials than PathTime-VLA under BC + DAgger at fixed 1×.

    Source: §VI · p. 7

  • Takeaways & Limitations

    The results support path-time factorization as a practical interface for reliable adaptation and efficient execution.

    Source: §VI · p. 7

    The study covers one platform, one teleoperation interface, and three tasks; timing also lacks dynamics-aware modeling of velocity-dependent effects, friction, or impact beyond demonstrated support.

    Source: §V · p. 7

Abstract

from arXiv · show

Vision-Language-Action (VLA) policies typically predict actions at fixed time intervals, coupling the route a robot follows with its execution pace. This coupling complicates adaptation from teleoperation: useful geometric guidance comes with timing shaped by interface delays and operator behavior. Our key insight is to bring the path-time parameterization of classical motion planning into the learned action representation of a VLA. We introduce PathTime-VLA, which represents motion as a progress-indexed interaction path $X(s)$ and a positive interval-time profile. The latter defines a monotone time law $t(s)$, yielding controller commands $X(s(t))$. 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, Speed-DQN 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. Across three tasks, the complete method achieves $58/60$ successes versus $57/60$ for PathTime-VLA under BC + DAgger at fixed $1\times$, with approximately $39$-$52\%$ shorter mean completion times over successful trials.

I. INTRODUCTION

Fixed-time VLA actions couple path geometry with execution pace, while teleoperated timing can reflect interface delays and operator behavior. PathTime-VLA separates path and timing so they can be post-trained with distinct signals.

I. INTRODUCTION

  • Fixed-interval action chunks entangle execution timing with the geometric sequence used as supervision, complicating timing improvement.Teleoperated timing may reflect network latency, interface limits, and operator behavior rather than an efficient autonomous schedule.Source: §I · p. 1
  • Demonstrations and DAgger establish a target-domain prior, Speed-DQN learns chunk-level pacing with the path generator fixed, and Path-AWR refines paths from rollout outcomes.The stages use supervision suited to path generation and execution timing, respectively.Source: §I · p. 2
  • PathTime-VLA separates a progress-indexed interaction path from an interval-time profile, exposing distinct post-training interfaces for geometry and timing.Composing the path with its timing recovers controller-rate actions; different schedules can leave the geometric target unchanged.Source: §I · pp. 1–2
  • The study spans three manipulation tasks and includes teleoperated demonstrations, DAgger interventions, online DQN rollouts, and 600 real-robot evaluation trials.Source: §I · p. 2

II. RELATED WORK

The related work spans VLA action generation, classical path-time planning, and interactive post-training. PathTime-VLA combines these threads by separating timing choices from path generation and assigning them distinct learning signals.

A. Large-Scale VLA Pretraining and Action Generation

B. Path-Time Parameterization and Speed Control

  • Classical planning retimes prescribed paths, while PathTime-VLA pairs a progress-indexed path with interval traversal times so retiming leaves the geometric target unchanged.Related learned policies address execution cadence or physical speed through high-frequency execution, asynchronous chunks, speed modulation, or time-indexed outputs conditioned on desired speed.Source: B. Path-Time Parameterization and… · p. 2

C. Interactive Imitation and Reinforcement Post-Training

  • DAgger methods expand coverage around failures, but timestamped imitation inherits operator- and interface-dependent timing; PathTime-VLA combines discrete timing-value learning with rollout-based path refinement.The refinement is isolated to path generation rather than direct diffusion-policy or general offline-to-online actor updates.Source: C. Interactive Imitation and… · p. 2

III. METHOD

  • PathTime-VLA reuses geometric guidance across execution rates, matching dense corrective paths from interventions with task outcomes that evaluate timing choices and realized paths.Its post-training combines supervised aggregation, low-dimensional timing reinforcement learning, and advantage-weighted path refinement.Source: §III · p. 2

A. Path-Time Action Representation

PathTime-VLA represents a chunk using path samples and their traversal times rather than a fixed-time action sequence. This allows timing changes to retime a path without redefining its geometric target.

1) From Time-Indexed Chunks to Path-Time Coordinates:

2) Constructing Path and Timing Targets:

The target representation separates fixed-progress interaction paths from interval timing, then composes them into controller-rate commands. Timing variants preserve a shared geometric target, while the action expert assigns phase-dependent durations to the path.

2) Constructing Path and Timing Targets:

  • Fixed-progress path knots encode pose and gripper commands, while valid interval times define a monotone time law for controller-rate commands.Force and torque are retained as auxiliary channels; padded endpoint intervals are excluded from timing supervision and time-law construction.Source: 2) Constructing Path and Timing Targets · p. 3

3) Timing Variants on a Shared Path:

  • Timing variants divide valid durations by a multiplier and scale the rate condition while keeping the geometric target unchanged.At deployment, the path and multiplier are selected anew from the latest observation, so the invariance applies to a retimed chunk target rather than a globally fixed route.Source: 3) Timing Variants on a Shared Path · p. 3

4) Action Expert Architecture:

  • The PathHead predicts paths without a speed multiplier, while the TimeHead assigns phase-dependent interval times to realize the requested global pace.The TimeHead can vary timing across free-space motion, contact, engagement, and release; stop-gradient prevents timing loss from changing path geometry.Source: 4) Action Expert Architecture · p. 3

B. Factorized Post-Training

Path-Time factorization gives geometry and timing distinct post-training interfaces. The pipeline establishes a supervised prior, explores timing choices on the robot, and uses rollout experience to refine path generation.

B. Factorized Post-Training

  • Demonstrations and interventions provide dense corrective targets, while task outcomes evaluate resulting executions.The stages pair supervised adaptation with low-dimensional reinforcement learning for timing and advantage-weighted diffusion training for path generation.Source: B. Factorized Post-Training · p. 3
  • Supervised adaptation establishes the action expert before Speed-DQN explores timing choices on the robot and Path-AWR uses that experience to refine path generation.The path-time representation exposes alternative timing choices for a shared path target.Source: B. Factorized Post-Training · p. 3, Figure 3 · p. 4

1) Supervised Adaptation with Demonstrations and DAgger:

The training procedure first builds a target-domain action-expert prior from demonstrations and DAgger corrections. It then explores execution multipliers with Speed-DQN and uses outcome-labeled paths for a conservative Path-AWR update.

1) Supervised Adaptation with Demonstrations and DAgger:

  • Demonstrations train the PathHead on interaction paths, the VelocityHead on demonstrated base rates, and the TimeHead on durations under augmented rate conditions.DAgger corrections add local paths and recovery trajectories from failure-prone states; their timing supervises base rate but is not treated as uniquely optimal.Source: 1) Supervised Adaptation with… · p. 4

Speed-DQN: Learning Execution Multipliers

  • Speed-DQN explores a small discrete timing space while keeping the demonstrated path prior fixed, avoiding direct reinforcement learning over high-dimensional diffusion paths.It selects execution multipliers per chunk without jointly perturbing path geometry.Source: 2) Speed-DQN: Learning Execution… · p. 4
  • The reward combines terminal task feedback with a bounded preference for faster bins, with terminal feedback dominant so speed is optimized within task-success reliability.Slower bins incur an opportunity cost without positive per-chunk reward that could favor longer episodes.Source: 2) Speed-DQN: Learning Execution… · p. 4

Path-AWR: Learning from Executed Paths

  • Path-AWR uses execution-aligned, outcome-labeled path samples for a conservative update, reinforcing higher-advantage online paths while anchoring the generator to BC + DAgger paths.Safe-sample advantages map to monotone quantile weights, unsafe samples receive zero, and only the PathHead is updated.Source: 3) Path-AWR: Learning from Executed… · pp. 4–5

4) Training Summary:

The training pipeline hands off from supervised adaptation to online speed learning and path refinement. It constructs path-time targets, collects interventions, then uses execution-aligned replay to update only the PathHead.

4) Training Summary:

  • The pipeline constructs fixed-progress paths, interval times, and timing variants from demonstrations, then trains the PathHead, VelocityHead, and TimeHead with supervised loss.It rolls out at fixed 1×, collects DAgger interventions, and fine-tunes on the combined demonstration and intervention data.Source: 4) Training Summary · p. 5
  • After supervised fine-tuning, the pipeline freezes the VLA during online Speed-DQN training and collects execution-aligned path replay.It fits a path critic and state-value baseline on terminal path returns, then converts path advantages to safe quantile weights.Source: 4) Training Summary · p. 5
  • Path-AWR unfreezes and updates only the PathHead using the weighted replay and supervised objective.The algorithm summary specifies that the PathHead alone is updated with Eq. (10).Source: 4) Training Summary · p. 5

A. Real-Robot Tasks

The evaluation spans three manipulation tasks, from precise contact and articulated-object interaction to a longer multi-stage sequence.

A. Real-Robot Tasks

  • The three tasks cover precise contact, articulated-object interaction, and a longer multi-stage sequence.Peg insertion requires at least 1 cm of insertion; microwave opening requires the door to exceed 30°; tape placement requires closing the drawer to a gap of at most 5 cm.Source: A. Real-Robot Tasks · p. 5

B. Platform and Evaluation Protocol

The protocol matches data and evaluation conditions across methods; before reinforcement learning, PathTime-VLA outperforms the coupled representation, while fixed-rate effects vary by task and supervision.

B. Platform and Evaluation Protocol

  • The study uses shared robot and control settings, matched randomized trials, and identical demonstration data, initialization, and training budgets for the representation comparison.Each condition receives 20 trials per task; failures include safety stops and timeouts.Source: B. Platform and Evaluation Protocol · p. 5

C. Effect of the Action Representation

  • PathTime-VLA achieves 43/60 versus 40/60 successes under BC and 57/60 versus 51/60 under BC + DAgger against the coupled representation at fixed 1×.Mean completion time over successful trials is lower for PathTime-VLA in all six task–supervision comparisons.Source: C. Effect of the Action Representation · p. 5, Table I · p. 6

D. Execution-Rate Control with Fixed Multipliers

E. From Human Corrections to Outcome-Guided Refinement

Human corrections, execution feedback, and rollout outcomes support staged refinement of timing and paths. The complete method achieves similar aggregate success to BC + DAgger at fixed 1× with shorter successful-trial times, within a study limited to one platform, interface, and three tasks.

E. From Human Corrections to Outcome-Guided Refinement

  • Speed-DQN reduces successful-trial times but changes aggregate success from 57/60 to 55/60; Path-AWR raises microwave-opening success from 17/20 to 20/20.Path-AWR retains 20/20 insertion successes and 18/20 tape-and-drawer successes.Source: E. From Human Corrections to… · p. 6

VI. CONCLUSION

  • Across 600 real-robot evaluation trials on three manipulation tasks, the complete method achieves 58/60 successes and, versus PathTime-VLA with BC + DAgger at fixed 1×, reduces successful-trial mean completion time by approximately 39–52%.It also achieves the highest throughput proxy on each task among the evaluated fixed-rate and reinforcement post-training variants.Source: §VI · p. 7

V. LIMITATIONS

  • The study covers one platform, one teleoperation interface, and three tasks, so broader operators, embodiments, and task families remain future evaluation targets.Fixed-progress discretization trades geometric detail against horizon, and synthetic scaling cannot reproduce velocity-dependent dynamics, friction, or impact outside demonstrated support.Source: §V · p. 7

VI. CONCLUSION

  • PathTime-VLA separates the progress-indexed path from its interval-time profile, assigning DAgger, Speed-DQN, and Path-AWR distinct learning signals for the target-domain prior, timing, and path refinement.The composed path and timing are still evaluated together through task outcomes.Source: §VI · p. 7

Related papers

Loading 2610.11771v1…