Generalizing to long-horizon manipulation tasks in a zero-shot setting remains a central challenge in robotics. Current multimodal foundation based approaches, despite their capabilities, typically fail to decompose high-level commands into executable action sequences from static visual input alone. To address this challenge, we introduce Super-Mimic, a hierarchical framework that enables zero-shot robotic imitation by directly inferring procedural intent from unscripted human demonstration videos. Our framework is composed of two sequential modules. First, a Human Intent Translator (HIT) parses the input video using multimodal reasoning to produce a sequence of language-grounded subtasks. These subtasks then condition a Future Dynamics Predictor (FDP), which employs a generative model that synthesizes a physically plausible video rollout for each step. The resulting visual trajectories are dynamics-aware, explicitly modeling crucial object interactions and contact points to guide the low-level controller. We validate this approach through extensive experiments on a suite of long-horizon manipulation tasks, where Super-Mimic significantly outperforms state-of-the-art zero-shot methods by over 20%. These results establish that coupling video-driven intent parsing with prospective dynamics modeling is a highly effective strategy for developing general-purpose robotic systems.

We propose Super-Mimic, a hierarchical framework that combines the HIT module for task planning and the FDP module for subtask execution to tackle zero-shot long-horizon robotic manipulation.

The Human Intent Translator (HIT): This module parses complex human actions into transferable, language-guided robotic plans. It extracts key action frames from human videos and generates language-conditioned, transferable subtask plans through multimodal reasoning.

The Future Dynamics Predictor (FDP): This module employs video generation dynamics-aware visual trajectories that capture object interactions and contact constraints.

These imagined futures guide low-level control with enhanced robustness in cluttered, open-world environments.

We evaluate Super-Mimic on three challenging long-horizon scenarios, each comprising 5 to 8 subtasks:

• Meal Preparation: A 5-subtask sorting task requiring the robot to categorize fruits and snack boxes into different containers.
• Tidy Up the Desk: A 5-subtask task where the robot must use commonsense reasoning to sort trash from useful items into a bin and a basket.
• Irregular Traversal: An 8-subtask task with non-logical pick-and-place motions, designed to be difficult to describe with language.

Our experimental platform: a 7-DoF xArm7 arm and a third-view Orbbec RGB-D camera.

Our experiments demonstrate that Super-Mimic significantly outperforms state-of-the-art zero-shot methods by over 20% on these tasks.

We evaluate Super-Mimic on three challenging long-horizon scenarios and compare against state-of-the-art zero-shot methods ReKep and MOKA. The results show that Super-Mimic significantly outperforms the baselines by over 20%.

Key Results:

• Super-Mimic achieves 50% TSR on Meal Preparation, more than doubling the 20% TSR of the strongest baseline (MOKA).
• On Irregular Traversal, Super-Mimic achieves 20% TSR while baselines fail entirely.
• The framework demonstrates unique flexibility in Video + Constraints and Skill Generalization settings.