Strengthening Generative Robot Policies through Predictive World Modeling
Abstract
We present generative predictive control (GPC), a learning control framework that (i) clones a generative diffusion-based policy from expert demonstrations, (ii) trains a predictive action-conditioned world model from both expert demonstrations and random explorations, and (iii) synthesizes an online planner that ranks and optimizes the action proposals from (i) by looking ahead into the future using the world model from (ii). Crucially, we show that conditional video diffusion allows learning (near) physics-accurate visual world models and enable robust visual foresight. Focusing on planar pushing with rich contact and collision, we show GPC dominates behavior cloning across state-based and vision-based, simulated and real-world experiments.
World Model Prediction
Plain Push-T
World Model Prediction in GPC-RANK.
World Model Prediction in GPC-OPT.
Push-T collided with A
World Model Prediction in GPC-RANK.
World Model Prediction in GPC-OPT.
Push-T collided with A & R
World Model Prediction in GPC-RANK.
World Model Prediction in GPC-OPT.
Real-world Evaluation
Check all real-world evalutaion results (by clicking the titles)
Plain Push-T: Baseline (5 out of 10) vs. GPC-RANK (7 out of 10) vs. GPC-OPT (7 out of 10)
Baseline Test 0: Success.
GPC-RANK Test 0: Success.
GPC-OPT Test 0: Success.
Baseline Test 1: Failure.
GPC-RANK Test 1: Success.
GPC-OPT Test 1: Success.
Baseline Test 2: Success.
GPC-RANK Test 2: Success.
GPC-OPT Test 2: Success.
Baseline Test 3: Failure.
GPC-RANK Test 3: Failure.
GPC-OPT Test 3: Success.
Baseline Test 4: Failure.
GPC-RANK Test 4: Success.
GPC-OPT Test 4: Success.
Baseline Test 5: Failure.
GPC-RANK Test 5: Success.
GPC-OPT Test 5: Failure.
Baseline Test 6: Success.
GPC-RANK Test 6: Failure.
GPC-OPT Test 6: Failure.
Baseline Test 7: Failure.
GPC-RANK Test 7: Failure.
GPC-OPT Test 7: Success.
Baseline Test 8: Success.
GPC-RANK Test 8: Success.
GPC-OPT Test 8: Success.
Baseline Test 9: Success.
GPC-RANK Test 9: Success.
GPC-OPT Test 9: Failure.
Push-T collided with A: Baseline (2 out of 5) vs. GPC-RANK (4 out of 5) vs. GPC-OPT (3 out of 5)
Baseline Test 0: Failure.
GPC-RANK Test 0: Success.
GPC-OPT Test 0: Success.
Baseline Test 1: Success.
GPC-RANK Test 1: Success.
GPC-OPT Test 1: Success.
Baseline Test 2: Failure.
GPC-RANK Test 2: Success.
GPC-OPT Test 2: Failure.
Baseline Test 3: Failure.
GPC-RANK Test 3: Failure.
GPC-OPT Test 3: Failure.
Baseline Test 4: Success.
GPC-RANK Test 4: Success.
GPC-OPT Test 4: Success.
Push-T collided with A & R: Baseline (2 out of 5) vs. GPC-RANK (3 out of 5) vs. GPC-OPT (4 out of 5)
Baseline Test 0: Success.
GPC-RANK Test 0: Success.
GPC-OPT Test 0: Success.
Baseline Test 1: Success.
GPC-RANK Test 1: Success.
GPC-OPT Test 1: Success.
Baseline Test 2: Failure.
GPC-RANK Test 2: Success.
GPC-OPT Test 2: Success.
Baseline Test 3: Failure.
GPC-RANK Test 3: Failure.
GPC-OPT Test 3: Failure.
Baseline Test 4: Failure.
GPC-RANK Test 4: Failure.
GPC-OPT Test 4: Success.
Push-T collided with R: Baseline (2 out of 3) vs. GPC-RANK (3 out of 3) vs. GPC-OPT (3 out of 3)
Baseline Test 0: Success.
GPC-RANK Test 0: Success.
GPC-OPT Test 0: Success.
Baseline Test 1: Success.
GPC-RANK Test 1: Success.
GPC-OPT Test 1: Success.
Baseline Test 2: Failure.
GPC-RANK Test 2: Success.
GPC-OPT Test 2: Success.
BibTeX
@article{qi2025gpc,
title={Strengthening Generative Robot Policies through Predictive World Modeling},
author={Qi, Han and Yin, Haocheng and Du, Yilun and Yang, Heng},
journal={arXiv preprint arXiv:2502.00622},
year={2025}
}