RePlan-VLA: Progress-Aware Replanning for Long-Horizon Vision-Language-Action Models
Organizations: Purdue University, West Lafayette, IN, USA · DEVCOM ARL, Adelphi, MD, USA
Abstract
Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic manipulation, yet their performance often degrades on long-horizon tasks that require multiple interdependent manipulations. Instead of burdening the VLA policies with long-horizon reasoning, memory, action generation, progress tracking, and error recovery, we argue for leveraging their short-horizon manipulation capabilities within an execution-aware framework. To this end, we introduce RePlan-VLA, a VLA-agnostic framework that translates a VLA's short-horizon capabilities into reliable long-horizon execution. RePlan-VLA explicitly monitors the robot's motion for progress evaluation, and adaptively revises the future plan based on observed outcomes. We preserve the underlying VLA policy entirely, requiring no additional fine-tuning. Experiments on RoboCasa365 and CALVIN benchmarks show an improvement of 21.9% and 10.4% in success rate over the next strongest baseline. We also validate our approach for real-world deployment on a Franka Research 3 robot.
Figures & tables
| Approach | Success Rate (%) | ||
|---|---|---|---|
| Seen | Unseen | Avg. | |
| GR T-N1.6 [ 23 ] | 10.0 | 0.00 | 5.00 |
| [ 11 ] | 18.8 | 0.00 | 9.40 |
| RLDX-1 [ 12 ] | 16.3 | 6.30 | 11.3 |
| Xiaomi Robotics-1 [ 29 ] | 36.3 | 27.5 | 31.9 |
| Awomo-0.5I [ 27 ] | 55.0 | 26.3 | 40.6 |
| Approach | 1 | 2 | 3 | 4 | 5 | Avg. |
|---|---|---|---|---|---|---|
| [ 11 ] | 80.0 | 42.0 | 20.0 | 16.0 | 10.0 | 1.68 |
| FLOWER [ 26 ] | 94.0 | 68.0 | 44.0 | 42.0 | 38.0 | 2.86 |
| MDT [ 24 ] | 96.0 | 72.0 | 48.0 | 46.0 | 40.0 | 3.02 |
| XR-0 [ 4 ] | 90.0 | 78.0 | 66.0 | 64.0 | 60.0 | 3.58 |
| MoDe [ 25 ] | 86.0 | 62.0 | 46.0 | 36.0 | 30.0 | 2.60 |
| ForeACT [ 36 ] | 62.0 | 22.0 | 6.00 | 4.00 | 00.0 | 0.94 |
| Success Rate (%) | |||
| Configuration | Seen | Unseen | Avg. |
| Visual Grounding | |||
| SAM3 Segmentation Masks | |||
| Grounding DINO Bounding Boxes | |||
| Object Labels Only | |||
| Without Object Annotations | |||
| Metric | Seen | Unseen |
|---|---|---|
| Runtime Efficiency | ||
| Progress Estimation (s/call) | 2.31 | 2.39 |
| Re-planning (s/call) | 3.56 | 2.69 |
| Instruction Verification (s/call) | 0.23 | 0.21 |
| Object Detection (s/call) | 0.57 | 0.51 |
| Re-planning Behavior | ||
Long-Horizon Robotic Manipulation
Long-horizon robotic manipulation studies how robots plan, remember, and recover across extended sequences of physical tasks.
See all 466 papers