JEPA Policy: Diffusion-Free Imitation Learning via Paired Action and Future Representation Prediction
[Paper Quick Read] Robots Learn Actions and 'Future States' Simultaneously: High-Speed Imitation Learning with JEPA Policy
JEPA Policy improves imitation learning by predicting both action chunks and future state representations simultaneously within a single Transformer. It achieves an average success rate of 83.0% across nine simulation tasks, up from 77.4%, with minimal latency increase. The method avoids generating detailed future videos, instead focusing on latent representations that capture task-relevant outcomes.
The paper introduces JEPA Policy, a method for imitation learning that jointly predicts action chunks and future state representations, achieving a 5.6 percentage point improvement in success rate across nine simulation tasks without using diffusion models.
Source: Robotics — Robot learning news discovery · 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 S12
Reported numbers
latency increase
0.29
View original evidence
as a pair. It allows both tokens to see each other within a single Transformer and refines coarse predictions through two forward passes. Without using diffusion's iterative denoising, it increased the average success rate from 77.4% to 83.0% across nine simulation tasks while keeping the latency increase relative to MIP at 0.29ms. Title: JEPA Policy: Diffusion-Free Imitation Learning via Paired Action and Future Rep
Open source S3latency
13.2
View original evidence
obot domains. Diagnosing future anomalies in 13.2ms Figure 6: Comparison of decision latency between JEPA Policy and Diffusion Policy Measured on the same accelerator with batch size 1 on Tool Hang, MIP was 12.9ms and JEPA Policy was 13.2ms. The additional cost of the future branch is 0.29ms, and the parameter increase is 0.79M. A 100-step Diffusion Policy takes 439.5ms, while JEPA Policy can make decisions at approx
Open source S14latency
12.9
View original evidence
obot domains. Diagnosing future anomalies in 13.2ms Figure 6: Comparison of decision latency between JEPA Policy and Diffusion Policy Measured on the same accelerator with batch size 1 on Tool Hang, MIP was 12.9ms and JEPA Policy was 13.2ms. The additional cost of the future branch is 0.29ms, and the parameter increase is 0.79M. A 100-step Diffusion Policy takes 439.5ms, while JEPA Policy can make decisions at approx
Open source S14
- paper: Available Artifact link Open source S4
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
of success rates across 9 robot tasks For simulation evaluation, we used MokaMoka, MugMug, Square, Tool Hang, Transport, Coffee Preparation, Kitchen, Three Piece Assembly, and Hammer Cleanup. JEPA Policy outperformed action-only MIP in all 9 tasks, with the average improving by 5.6 points from 77.4% to 83.0%. The largest improvements were in Coffee Preparation, from 57.5% to 77.5%, and Tool Hang, from 65.8% to 75.0%.
Open source S12
It also outperformed the Diffusion Policy's 9-task average of 75.1%. Gains of 4.7 to 5.3 points remained not only for the best checkpoint but also for the average of all 30 evaluations and the average of the last 5 evaluations. The pooled ranking was the same for 5 physical robot tasks totaling 630 episodes. However, the number of tasks and seeds was limited, so these results do not guarantee generalization to all r
Open source S13
as a pair. It allows both tokens to see each other within a single Transformer and refines coarse predictions through two forward passes. Without using diffusion's iterative denoising, it increased the average success rate from 77.4% to 83.0% across nine simulation tasks while keeping the latency increase relative to MIP at 0.29ms. Title: JEPA Policy: Diffusion-Free Imitation Learning via Paired Action and Future Rep
Open source S3
obot domains. Diagnosing future anomalies in 13.2ms Figure 6: Comparison of decision latency between JEPA Policy and Diffusion Policy Measured on the same accelerator with batch size 1 on Tool Hang, MIP was 12.9ms and JEPA Policy was 13.2ms. The additional cost of the future branch is 0.29ms, and the parameter increase is 0.79M. A 100-step Diffusion Policy takes 439.5ms, while JEPA Policy can make decisions at approx
Open source S14
ay not allow information to reach the action side. The paper prepared Dual-independent and Dual-cross as controls. The former separates the two branches, while the latter uses cross-attention but does not share all layers in a single stack like JEPA Policy. Across 9 tasks, Dual-independent averaged 76.3% and Dual-cross 77.6%, whereas JEPA Policy achieved 83.0%. An ablation that cuts the path for the future-loss gradi
Open source S10
resentation Prediction URL: https://arxiv.org/abs/2609.09630 Affiliation: Anyverse Dynamics Authors: Jie Xu, Kangjin Yu, Ziyi Jin, Junjie Gao, Liqing Chen, Yixian Li, Shuai Tian, Zhongpu Xia Figure 1: Design comparison of Diffusion Policy, latent JEPA, ACT-JEPA, and JEPA Policy Background: Action labels lack the meaning of results Behavior cloning directly regresses expert actions at:t+H from observations ot. While t
Open source S4
Source：Robotics — Robot learning news discovery · note.com