FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation

Minh Duc Nguyen, Nghiem Tuong Diep, Nguyen Gia Binh, Trong-Bao Ho, Doanh Le Thien, Quang Tan Nguyen, Thien-Loc Ha, Tran Van Nhiem, Bao Thach, Tran Xuan Nhat, Tuan Anh Tran, Artur Habuda, Philip Lund Møller, Tran Nguyen Le, Daniel Sonntag, Mathias Niepert, Khoa D Doan, Vu N. Duong, Hung Ngo, Minh Nhat Vu, Duy Minh Ho Nguyen, An Thai Le, Vien Anh Ngo
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:92316-92349, 2026.

Abstract

Vision–Language–Action (VLA) models enable general-purpose robotic control via large-scale multimodal pretraining, yet their effectiveness under few-shot imitation learning remains limited. We conduct a systematic stress test of state-of-the-art VLA models and show that performance degrades sharply as demonstrations are reduced, revealing a key weakness of existing adaptation strategies. To address this, we introduce FOCA, a future-oriented conditioning framework for data-efficient VLA adaptation. FOCA combines explicit prediction of task-grounded future interaction embeddings with implicit alignment to future goal observations, enabling long-horizon reasoning in latent space without pixel-level prediction. This formulation naturally supports action-free co-training with synthetic videos from video world models and can be interpreted as learning a future-conditioned value-like representation. Extensive experiments demonstrate FOCA achieves 95.7% success with 20 demonstrations on LIBERO, improves 7–12% on RoboCasa, and delivers up to 26% absolute gains on real robots, establishing a new state of the art in few-shot VLA adaptation.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-nguyen26b, title = {{FOCA}: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation}, author = {Nguyen, Minh Duc and Diep, Nghiem Tuong and Binh, Nguyen Gia and Ho, Trong-Bao and Thien, Doanh Le and Nguyen, Quang Tan and Ha, Thien-Loc and Van Nhiem, Tran and Thach, Bao and Nhat, Tran Xuan and Tran, Tuan Anh and Habuda, Artur and M{\o}ller, Philip Lund and Nguyen Le, Tran and Sonntag, Daniel and Niepert, Mathias and Doan, Khoa D and Duong, Vu N. and Ngo, Hung and Vu, Minh Nhat and Nguyen, Duy Minh Ho and Le, An Thai and Ngo, Vien Anh}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {92316--92349}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/nguyen26b/nguyen26b.pdf}, url = {https://proceedings.mlr.press/v306/nguyen26b.html}, abstract = {Vision–Language–Action (VLA) models enable general-purpose robotic control via large-scale multimodal pretraining, yet their effectiveness under few-shot imitation learning remains limited. We conduct a systematic stress test of state-of-the-art VLA models and show that performance degrades sharply as demonstrations are reduced, revealing a key weakness of existing adaptation strategies. To address this, we introduce FOCA, a future-oriented conditioning framework for data-efficient VLA adaptation. FOCA combines explicit prediction of task-grounded future interaction embeddings with implicit alignment to future goal observations, enabling long-horizon reasoning in latent space without pixel-level prediction. This formulation naturally supports action-free co-training with synthetic videos from video world models and can be interpreted as learning a future-conditioned value-like representation. Extensive experiments demonstrate FOCA achieves 95.7% success with 20 demonstrations on LIBERO, improves 7–12% on RoboCasa, and delivers up to 26% absolute gains on real robots, establishing a new state of the art in few-shot VLA adaptation.} }
Endnote
%0 Conference Paper %T FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation %A Minh Duc Nguyen %A Nghiem Tuong Diep %A Nguyen Gia Binh %A Trong-Bao Ho %A Doanh Le Thien %A Quang Tan Nguyen %A Thien-Loc Ha %A Tran Van Nhiem %A Bao Thach %A Tran Xuan Nhat %A Tuan Anh Tran %A Artur Habuda %A Philip Lund Møller %A Tran Nguyen Le %A Daniel Sonntag %A Mathias Niepert %A Khoa D Doan %A Vu N. Duong %A Hung Ngo %A Minh Nhat Vu %A Duy Minh Ho Nguyen %A An Thai Le %A Vien Anh Ngo %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-nguyen26b %I PMLR %P 92316--92349 %U https://proceedings.mlr.press/v306/nguyen26b.html %V 306 %X Vision–Language–Action (VLA) models enable general-purpose robotic control via large-scale multimodal pretraining, yet their effectiveness under few-shot imitation learning remains limited. We conduct a systematic stress test of state-of-the-art VLA models and show that performance degrades sharply as demonstrations are reduced, revealing a key weakness of existing adaptation strategies. To address this, we introduce FOCA, a future-oriented conditioning framework for data-efficient VLA adaptation. FOCA combines explicit prediction of task-grounded future interaction embeddings with implicit alignment to future goal observations, enabling long-horizon reasoning in latent space without pixel-level prediction. This formulation naturally supports action-free co-training with synthetic videos from video world models and can be interpreted as learning a future-conditioned value-like representation. Extensive experiments demonstrate FOCA achieves 95.7% success with 20 demonstrations on LIBERO, improves 7–12% on RoboCasa, and delivers up to 26% absolute gains on real robots, establishing a new state of the art in few-shot VLA adaptation.
APA
Nguyen, M.D., Diep, N.T., Binh, N.G., Ho, T., Thien, D.L., Nguyen, Q.T., Ha, T., Van Nhiem, T., Thach, B., Nhat, T.X., Tran, T.A., Habuda, A., Møller, P.L., Nguyen Le, T., Sonntag, D., Niepert, M., Doan, K.D., Duong, V.N., Ngo, H., Vu, M.N., Nguyen, D.M.H., Le, A.T. & Ngo, V.A.. (2026). FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:92316-92349 Available from https://proceedings.mlr.press/v306/nguyen26b.html.

Related Material