VLA-MOPD: On-Policy Distillation for Multi-Task VLA Policy Consolidation
VLA-MOPD: On-Policy Distillation for Multi-Task VLA Policy Consolidation
This work introduces VLA-MOPD, a method for consolidating multi-task Vision-Language-Action policies through on-policy distillation. The approach uses frozen instruction-routed teachers queried at each visited observation and intermediate ODE state to supervise action chunks via masked local matching loss, with an optional reference-policy term to retain skills across tasks. Starting from one trajectory per LIBERO task, the method raises mean success-once from 75.4% to 98.9%, compared with 98.15% for endpoint action distillation. It improves mean RoboTwin 2.0 success from 73.0% to 89.3%; on a Piper robot, two consolidation rounds improve average success from 70.0% to 90.0%. In a near-matched Spatial-adaptation study, the reference term limits the Object suite decrease to 3.0 points versus 13.2 without it.
The paper proposes VLA-MOPD, a distillation protocol that consolidates multi-task Vision-Language-Action policies by querying frozen teachers at intermediate ODE states and using masked local matching loss, reporting success rate improvements on LIBERO (75.4% to 98.9%) and RoboTwin 2.0 (73.0% to 89.3%) under the stated experimental conditions.
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
Object suite decrease with reference matching
3
View original evidence
an RoboTwin 2.0 success from 73.0% to 89.3%; on a Piper robot, two consolidation rounds improve average success from 70.0% to 90.0%. In a near-matched Spatial-adaptation study, reference matching limits the Object suite decrease to 3.0 points, versus 13.2 without the reference term.
Open source S5Object suite decrease without reference term
13.2
View original evidence
an RoboTwin 2.0 success from 73.0% to 89.3%; on a Piper robot, two consolidation rounds improve average success from 70.0% to 90.0%. In a near-matched Spatial-adaptation study, reference matching limits the Object suite decrease to 3.0 points, versus 13.2 without the reference term.
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
servation and intermediate ordinary differential equation (ODE) state. Detached per-flow-step transition targets supervise valid action chunks through a masked local matching loss; an optional reference-policy term supports retention on other tasks. Starting from one trajectory per LIBERO task, VLA-MOPD raises mean success-once from 75.4% to 98.9%, compared with 98.15% for endpoint action distillation. It improves me
Open source S4
an RoboTwin 2.0 success from 73.0% to 89.3%; on a Piper robot, two consolidation rounds improve average success from 70.0% to 90.0%. In a near-matched Spatial-adaptation study, reference matching limits the Object suite decrease to 3.0 points, versus 13.2 without the reference term.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org