Amazon's ResMimic Teaches Humanoids to Handle Objects by Adding Precision to General Motion

Amazon’s Frontier AI & Robotics (FAR) team has unveiled ResMimic, a new framework designed to bridge the gap between general human motion and precise, real-world object manipulation for humanoid robots. In a new research paper, the team demonstrates how a two-stage "residual learning" approach allows a Unitree G1 humanoid to learn complex loco-manipulation tasks—like lifting heavy boxes and irregularly shaped chairs—with greater efficiency and success than training from scratch.

The work tackles a persistent challenge in robotics: while policies can be trained to make a robot imitate human movements, this general skill often fails when precise interaction with the world is required. Directly copying a human's actions onto a robot with a different body shape and physical constraints often results in errors like hands passing through objects or losing grip.

ResMimic solves this by separating the problem into two parts. First, it uses a pre-trained General Motion Tracking (GMT) policy as a foundation. This base policy, trained on a massive 42-hour dataset of human-only motion, provides the robot with a robust, task-agnostic understanding of how to move like a person.

The key innovation comes in the second stage. Instead of retraining the entire policy for a specific task, ResMimic adds a small, lightweight "residual" policy on top. This secondary policy learns to make minor, corrective adjustments to the actions proposed by the base GMT policy. It learns the final bit of precision needed to correctly grasp, lift, and carry an object, effectively adding object-awareness to the robot's general movement skills.

From Coarse Imitation to Fine-Grained Skill

According to the paper, this residual learning approach is far more efficient than other methods. The core insight is that most of the required motion—balancing, stepping, reaching—is already captured by the general policy. The residual policy only needs to learn the small, task-specific difference, or "residual," required to succeed.

To help this process, the researchers developed several techniques to make training more stable, including a point-cloud-based reward for tracking objects and a "contact reward" that encourages the robot to use its whole body for support, much like a human would.

The results, demonstrated on a Unitree G1, are compelling. The robot is shown performing a range of difficult tasks that require both mobility and careful manipulation:

• Kneeling on one knee to lift a box.
• Lifting a 4.5 kg (9.9 lbs) box by using its torso and arms for support—exceeding the robot's wrist payload limit.
• Picking up and carrying heavy, awkwardly shaped chairs.
• Sitting down on a chair and standing back up.

In simulated tests, ResMimic achieved a 92.5% task success rate, dramatically outperforming policies trained from scratch (0% success) or fine-tuned from the base policy (7.5% success).

Building on a Foundation

ResMimic is the latest project from the same Amazon lab that recently developed OmniRetarget, a system for generating vast amounts of high-quality training data from a single human demonstration. The two projects appear complementary: OmniRetarget focuses on creating better data, while ResMimic offers a more efficient way to learn from that data.

Both projects utilize the Unitree G1, reinforcing the platform's growing role as a standard for cutting-edge robotics research due to its accessibility and capability. The research also aligns with Amazon's broader ambitions in humanoid robotics, as developing these foundational manipulation skills is a critical step for any future deployment in logistics. By building on a foundation of general skills, ResMimic represents another step toward creating humanoid robots that can learn useful, physical tasks in a scalable way, without requiring exhaustive, task-specific engineering for every new skill.