Coordinating Humanoid Loco-Manipulation through Action-Conditioned Latent Foresight
Coordinating Humanoid Loco-Manipulation through Action-Conditioned Latent Foresight
This research presents action-conditioned latent foresight, a latent world-action model (WAM) designed to evaluate proposed command sequences for humanoid robots before execution. The system uses an action generator to propose joint motion and hand commands based on vision, language, and robot state, while an endpoint predictor estimates visual features and progress toward subgoals. A joint temporal prior scores temporal consistency, and action consistency regularization ensures the predictor retains action-relevant information during training. On the SIMPLE simulation benchmark, the method achieves 89.7% average success across ten tasks, outperforming Psi-Zero (73.7%) and DiT4DiT (82.0%). In real-world tests on seven tasks, it reaches 85.7% success under standard conditions and 65.7% under unseen object configurations, compared to GR00T N1.7's 70.0% and 48.6%, respectively.
The paper introduces a latent world-action model for humanoid loco-manipulation that evaluates command sequences before execution, reporting 89.7% average success on ten simulation tasks in the SIMPLE benchmark and 85.7% on seven real-robot tasks under standard conditions.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
from the next observation. During training, action consistency regularization reconstructs the first demonstrated command from observed and predicted feature changes, encouraging the predictor to retain action-relevant information. On ten tasks in the humanoid loco-manipulation simulation benchmark SIMPLE, our method achieves 89.7% average success, compared with 73.7% for Psi-Zero and 82.0% for DiT4DiT. On seven rea
Open source S4
l-robot tasks, it achieves 85.7% under standard conditions and 65.7% under unseen object configurations, while GR00T N1.7 achieves 70.0% and 48.6%, respectively. Project page: https://ju6276.github.io/Action-Grounded-Foresight/
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org