ForeDiffusion: Foresight-Conditioned Diffusion Policy via Future View Construction for Robot Manipulation

TL;DR

ForeDiffusion addresses the fragility of diffusion-based visuomotor policies in long-horizon manipulation by conditioning each denoising step on a compact predicted future view constructed from near-term observations. By injecting the future view at a mid-stage within the diffusion process and training with a dual objective that couples denoising fidelity with future-structure consistency, the method mitigates error accumulation and enhances forward-looking planning. Empirically, it achieves an average success rate of $80.6\%$ on Adroit and MetaWorld, outperforming strong baselines by up to $23\%$ on complex tasks, with improvements in learning efficiency and robustness across task difficulty. These results suggest foresight-conditioned diffusion as a practical, data-efficient paradigm for reliable long-horizon robotic manipulation.

Abstract

Diffusion strategies have advanced visual motor control by progressively denoising high-dimensional action sequences, providing a promising method for robot manipulation. However, as task complexity increases, the success rate of existing baseline models decreases considerably. Analysis indicates that current diffusion strategies are confronted with two limitations. First, these strategies only rely on short-term observations as conditions. Second, the training objective remains limited to a single denoising loss, which leads to error accumulation and causes grasping deviations. To address these limitations, this paper proposes Foresight-Conditioned Diffusion (ForeDiffusion), by injecting the predicted future view representation into the diffusion process. As a result, the policy is guided to be forward-looking, enabling it to correct trajectory deviations. Following this design, ForeDiffusion employs a dual loss mechanism, combining the traditional denoising loss and the consistency loss of future observations, to achieve the unified optimization. Extensive evaluation on the Adroit suite and the MetaWorld benchmark demonstrates that ForeDiffusion achieves an average success rate of 80% for the overall task, significantly outperforming the existing mainstream diffusion methods by 23% in complex tasks, while maintaining more stable performance across the entire tasks.

Figures (7)

• Figure 1: Overview of ForeDiffusion. (a) Diagram of future view-guided action generation; (b) ForeDiffusion achieves the highest average success rate across all task types; (c) ForeDiffusion shifts task counts toward the higher success rate bins across all tasks chi2023diffusionpolicyZe2024DP3lu2025manicmrealtime3ddiffusionjia2024scoredistributionmatchingpolicy.
• Figure 2: Architecture of ForeDiffusion. The perception module fuses RGB-D and proprioceptive inputs into 3D latent representations; an observation encoder outputs a global condition $G$ and a future condition $\hat{G}$. These condition guide a $K$-step reverse-diffusion process that denoises a noise-perturbed action trajectory into an executable sequence $A_0$, with joint construction and behavioral losses enforcing accurate future prediction and expert-level control.
• Figure 3: Architecture of Foresight-Diffusion. A ResNet encoder–decoder injects 384 / 512-dimensional context vectors $G$ and $\hat{G}$ across all sampling stages to denoise the action token $A_{k}^{t}$ into $\hat{A}_{k}^{t}$, with anticipated future observations.
• Figure 4: Schematic of Future View Injection. Normal denoising performs the standard early‑phase diffusion, whereas foresight conditioning injects future‑view information, allowing the model to anticipate consecutive outcomes and generate more stable, goal‑aligned action sequences.
• Figure 5: Learning efficiency. Compared to DP3, ForeDiffusion shows higher stability, learning efficiency, and success rates.