TOPS improves robot task success with task-specific flow paths
TOPS: Reshaping Flow Source and Transport for Multi-Task Robot Manipulation · arXiv:2610.36694
In plain words
A robot policy can be asked to do many different tasks, but a standard flow-based policy starts each task from the same kind of random action and follows the same path toward its answer.
TOPS gives each known task its own starting distribution and adjusts the path toward demonstrated actions. In a simulated test of ten robot tasks, it had higher average success than standard flow matching, though the comparison also used different training continuation and action-generation steps.
The research
A shared robot policy has to produce different motions for different instructions, yet conventional flow policies begin from the same distribution and follow the same path for every task.
Can adapting where action generation starts and how it moves toward demonstrations help the policy complete more tasks?
TOPS fits a task-specific starting distribution and adjusts the path toward demonstrated actions, then selects both from the robot instruction.
To build these choices, the authors reverse-map demonstrated action chunks through a warm-up policy, fit a multimodal starting distribution for each task, and adjust interpolation where demonstrated actions vary more than starting samples.
In LIBERO-10 simulation, TOPS averaged 69.67% task success across three training runs, versus 53.44% for Standard Flow Matching; it used one generation step versus ten.
The result supports adapting the generation process in this tested setting. A source-only comparison also favors adding scheduling and its matching solver together, but does not isolate either one.
The LIBERO-10 comparison also changes TOPS’s continuation training and inference solver, and the paper does not specify whether the Standard FM arm starts from the same checkpoint. The result supports the tested setup, not a gain on every task or an isolated effect of one component.
What the evidence shows
- The shared-policy challenge
TOPS is a method for language-conditioned multi-task flow policies. The question is whether a shared robot policy can better handle tasks with different action patterns if generation starts from task-specific samples and follows a task-specific path.
Conventional flow policies use one shared Gaussian starting distribution and a common linear path, leaving the policy to recover task differences from the same starting setup.
- Looking back from actions
To inspect task differences, the authors ran a trained warm-up policy backward from expert action chunks, under the original observations and instructions. In the LIBERO-10 simulation, which tests robot manipulation across ten tasks, a visualization showed partially overlapping task groups and subgroups within tasks.
This projection suggests structure in the reverse-mapped samples, but it is not itself a measure of task success.
- Building task-specific paths
TOPS fits a multimodal starting distribution to each task’s reverse-mapped samples and compares that distribution with the variation in demonstrated actions. It changes the path most in directions where demonstrations vary more than the starting samples, moving slowly toward demonstrations early and faster later.
The sources and schedules are built offline, kept fixed during continuation training, and selected using the instruction.
- What the comparisons show
In simulated LIBERO-10, success means completing a manipulation task. Across three independent training runs, TOPS averaged 69.67% success versus 53.44% for Standard Flow Matching.
TOPS used one schedule-aware generation step and Standard Flow Matching used ten; both shared a policy framework and 600-epoch budget. A source-only control also supports a joint contribution from scheduling and its matching solver, though it cannot separate their effects.
- Evidence in a larger policy
The authors also applied TOPS to the larger π0.5 policy in LIBERO-Plus, which tests task completion under changed conditions. Across six reported categories, weighted by the number of variants and excluding camera changes where all methods scored zero, success rose from 64.4% for the baseline to 69.0% with TOPS.
These results extend the evidence to another policy and setting, while leaving unseen task identities untested.
Deep
- Original paper authors
- Tuo An, Chuhao Zhou, Jiahua Dong, Jingliang Li, Jiaqi Bai
- Submitted to arXiv
- License
- arXiv non-exclusive