PINT: A Unified Model with Shared
Predictive Intent for Robot Manipulation

Xinyu Guo1,2,3,*, Chao Jing4,3,*, Li Yi5, Xingyu Chen3,†,‡, Zhengxing Wu1,†

1Institute of Automation, Chinese Academy of Sciences 2University of Chinese Academy of Sciences
3Zhongguancun Academy 4Fudan University 5Tsinghua University

*Equal contribution †Corresponding authors ‡Project lead

guoxinyu2024@ia.ac.cn

PINT overview: shared Predictive Intent connects specialized experts, with two-stage training and simulation and real-world results.

Multiple experts, one intent. PINT coordinates semantic reasoning, world modeling, and action generation through shared chunk-level Predictive Intent while preserving expert-specific pathways.

Abstract

Generalist robot manipulation can benefit from pretrained vision-language models (VLMs) for semantic reasoning and video generation models (VGMs) for world modeling. However, existing unified models that directly fuse them through shared attention at multiple layers can make it harder to draw on their specialized capabilities. We introduce PINT, a unified model that coordinates semantic reasoning, world modeling, and action generation through a shared chunk-level Predictive Intent while preserving expert-specific pathways. PINT inserts learnable Intent Tokens into the VLM sequence to form this representation, conditioning both the VGM and an Action Expert (AE). It replaces only the VGM's prompt conditioning and retains the VGM at inference, where the AE reads its representations through one-way attention. Rather than prescribing what Predictive Intent should encode, PINT learns it jointly from the native prediction objectives of the VLM, VGM, and AE, which provide complementary views of the same upcoming task execution. Training uses two stages: Stage 1 freezes the VGM so that future video prediction shapes Predictive Intent with task-relevant future dynamics, while Stage 2 fixes the learned Intent and adapts the VGM and AE. Without embodied pretraining, PINT achieves strong in-distribution (ID) performance and substantially improves out-of-distribution (OOD) generalization, reaching 15.5% on RoboTwin 2.0-Hard compared with 1.8% for Fast-WAM and 3.2% for StarVLA, and 79.5% on LIBERO-Plus compared with 51.5% and 75.0%, respectively. Across five real-world bimanual tasks, PINT reaches 68% ID and 24% OOD success, compared with 35%/9% for pretrained Motus and 67%/35% for pretrained π0.5.

Method Overview

PINT architecture showing the VLM Intent Tokens, VGM, Action Expert, and one-way attention design.

PINT architecture and attention design. Learnable Intent Tokens form Predictive Intent, which conditions both the VGM and AE. The AE receives layer-wise current-frame keys and values from the VGM. The autoregressive action-token suffix is used only during training.

Future Video Prediction

ID and OOD future video predictions from the VGM trained in Fast-WAM and PINT after Stage 1 and Stage 2.

Future prediction across training stages. Predictions from the VGM trained in Fast-WAM and PINT after Stage 1 and Stage 2 use the same ID and OOD inputs. Stage 1 learns Predictive Intent with the VGM frozen; Stage 2 adapts the VGM to refine generation.

More Future Video Predictions

Closed-Loop Predictions

Expanded ID closed-loop predictions for placing a container on a plate, comparing Fast-WAM, PINT Stage 1, and PINT Stage 2.
ID
Expanded OOD closed-loop predictions for placing a container on a plate, comparing Fast-WAM, PINT Stage 1, and PINT Stage 2.
OOD

Expanded closed-loop examples for placing a container on a plate. Rows compare the VGM from our Fast-WAM reproduction with PINT after Stage 1 and Stage 2.

Open-Loop Predictions

Open-loop predictions, part I: ID examples on the left and OOD examples on the right, with ground truth, Fast-WAM, PINT Stage 1, and PINT Stage 2 rows. Open-loop predictions, part II, using the same ground-truth and model comparison format.

Open-loop future-video predictions on RoboTwin 2.0 demonstrations. Left and right panels show ID and OOD examples. Rows show ground truth, the VGM trained in Fast-WAM, PINT after Stage 1, and PINT after Stage 2. Columns show the main-camera view at t+8, t+16, t+24, and t+32.

Experimental Overview

PINT evaluation settings on LIBERO, LIBERO-Plus, RoboTwin 2.0, and five real-world bimanual tasks.

We evaluate PINT on LIBERO, LIBERO-Plus, RoboTwin 2.0, and five real-world bimanual tasks under ID and OOD settings.

Real-World Evaluation Videos

Representative successful PINT executions under in-distribution (ID) and out-of-distribution (OOD) settings.

Turn On Light (2×)

ID
OOD

Stack Cups (5×)

ID
OOD

Place Pens (5×)

ID
OOD

Store Block (5×)

ID
OOD

Fold Towel (7×)

ID
OOD

Simulation Results

LIBERO and LIBERO-Plus

Paper Table 1: full LIBERO and LIBERO-Plus results. PINT achieves 98.0% and 79.5%; PINT-VA achieves 98.2% and 81.0%.

RoboTwin 2.0

Paper Table 2: RoboTwin 2.0 Easy and Hard results. PINT achieves 68.6% and 15.5%; PINT-VA achieves 71.9% and 18.1%.

Real-World Results

Paper Table 3: real-world results on Turn On Light, Stack Cups, Place Pens, Store Block, and Fold Towel. PINT averages 68% ID and 24% OOD success.

Ablation Study

Paper Table 4: ablations of Intent design, supervision and expert interaction, two-stage training, and Intent capacity.

Visualizing Predictive Intent

Action chunks projected with t-SNE, colored by a PCA projection of their corresponding Predictive Intent, with representative manipulation behaviors.

Ground-truth action chunks are projected into a two-dimensional t-SNE space. Each point is colored using a three-dimensional PCA projection of its corresponding Predictive Intent. Highlighted regions show representative manipulation behaviors.

BibTeX

@article{guo2026pint,
  title={{PINT}: A Unified Model with Shared Predictive Intent for Robot Manipulation},
  author={Xinyu Guo and Chao Jing and Li Yi and Xingyu Chen and Zhengxing Wu},
  journal={arXiv preprint arXiv:2610.16355},
  year={2026},
  url={https://arxiv.org/abs/2610.16355}
}