Open-vocabulary mobile manipulation (OVMM) requires a robot to carry out a natural-language instruction on a mobile platform, and the success of learned manipulation policies hinges on placing the base within the policy's training distribution. We focus on policy mobilization, the problem of predicting an \(SE(2)\) base pose from which a fixed pre-trained manipulation policy will succeed, given a single ego-centric RGB-D frame and a language instruction. Prior work estimates a per-task, instruction-agnostic success likelihood over base poses, which limits generalization across tasks and language-based disambiguation. To address these limitations, we propose MoFlo, a multi-task model that learns a language- and observation-conditioned transport from the robot's current base pose to a policy-compatible endpoint, trained by conditional flow matching with \(x_1\)-prediction. On five RoboCasa kitchen tasks, MoFlo achieves an 80% mean success rate, outperforming prior policy-mobilization baselines, and further extends to language-based disambiguation among multiple fixtures of the same category. In real-world experiments on a mobile manipulator, MoFlo places the base for successful manipulation from a single ego-centric observation, including cases that require selecting a language-specified target in an ambiguous scene, and outperforms a representative baseline.
Prior methods ignore the instruction and train one model per task. Asked to “close the right drawer” with two drawers open, they put the robot in front of both and cannot tell which one is meant. MoFlo takes the instruction as input, so one model handles all tasks and picks the right target.
MoFlo encodes one RGB-D frame with Point-BERT and the instruction with CLIP. A transformer head then moves the robot's current base pose \(x_0\) to a target pose \(\hat{x}_1\) where the manipulation policy is likely to succeed.
No 3D reconstruction or pose search: MoFlo takes 44 ms per decision, while Mobi-\(\pi\) takes about 250 s plus 30 minutes of scene reconstruction.
MoFlo reaches 80% mean success on five kitchen tasks, ahead of N2M (63%) and Mobi-\(\pi\) (49%). It is best on four of five tasks. On TurnOnStove, the policy only works within about 5 cm of the ideal pose, and N2M edges ahead.
Precision is what matters here. MoFlo lands 5.3 cm from the ideal pose on average; random placement lands 15.4 cm away and reaches only 30%.
New phrasings. With unseen wording (“slide the drawer in” instead of “close the drawer”), MoFlo still reaches 70%, versus 80% with trained wording.
New visual styles. Training on more kitchen layouts improves success on kitchens with unseen appearance.
With two open drawers, MoFlo reaches the named one 62.5% of the time (Mobi-\(\pi\): 12.5%). With three drawers, never seen in training, it reaches 55.6%. Manipulation success is lower (37.5% and 24.4%) because the policy, trained on single-drawer scenes, sometimes bumps neighboring drawers.
Pose head. With everything else fixed, our \(x_1\)-prediction head scores best (80%). The other heads land within 8 points, except score-only (38%).
| Configuration | No language | Language |
|---|---|---|
| N2M | 63.0 | – |
| Per-task | 71.0 | 77.0 |
| Multi-task | 69.0 | 80.0 |
Where the gain comes from. Our pose head lifts N2M from 63% to 71%. Adding language and sharing one model across tasks brings it to 80%.
Encoders. RGB-only encoders work almost as well as Point-BERT. For text, CLIP beats larger language-only encoders.
On a Toyota HSR, MoFlo beats Mobi-\(\pi\) on all three tasks: Pick, CloseDrawer, and PushShelf. The first two need the instruction to pick the right target.
On Pick, a fine-tuned \(\pi_{0.5}\) that drives its own base succeeds 2 of 10 times. MoFlo with a fixed policy succeeds 5 of 10.
Pick
Pick
CloseDrawer: “close the orange drawer”
PushShelf: “push the shelf to the person”
Videos are played at 2.5× speed.
@inproceedings{yashima2026moflo,
title = {MoFlo: Language-Conditioned Flow Matching for Policy Mobilization},
author = {Yashima, Daichi and Takagi, Yusuke and Seno, Koki and Suzuki, Rui and Sakata, Kirato and Kurita, Shuhei and Sugiura, Komei},
booktitle = {Conference on Robot Learning (CoRL)},
year = {2026}
}