Grasp-TAGS

Overview

Unlike conventional mobile-manipulation settings that assume a fixed grasp configuration, Grasp-TAGS considers multiple feasible grasp poses on the target object. Each pose induces a distinct robot–object interaction configuration and, consequently, different manipulation dynamics. The objective is to learn one manipulation policy that transports the object to a desired planar pose under different initial grasps, then select at inference time the grasp with the highest expected utility for completing the rearrangement task.

Stage 1: Hierarchical Policy Learning

Large-object rearrangement requires coordinated locomotion and manipulation over long horizons. Because directly learning a unified whole-body policy over all robot joints is difficult in a high-dimensional action space with complex robot–object interactions, Grasp-TAGS adopts a hierarchical controller: a pretrained low-level locomotion controller maintains stable base motion, while a high-level task policy generates commands for object rearrangement.

The high-level command specifies base linear velocity, base yaw rate, and six arm-joint displacements. Arm commands are executed by PD controllers to hold the object, while the velocity commands are processed by the fixed locomotion policy to maintain stability and track the desired base motion. The high-level controller is trained with PPO in RSL-RL, while the pretrained low-level policy remains fixed.

Stage 2: Grasp-Conditioned Loco-manipulation Policy Training

Rather than training an independent high-level policy for every grasp, Grasp-TAGS trains one PPO policy across multiple initial grasp configurations. At the start of each episode, the environment samples a predefined feasible grasp pose; after establishing the grasp, the robot transports the object to a target planar pose. The initial grasp is implicitly represented by the arm joint angles, allowing the policy to learn behaviors across both grasp configurations and rearrangement objectives.

This unified policy captures the diverse interaction dynamics induced by different robot–object contact configurations without separate training procedures. Its value function also provides a common measure of expected long-horizon manipulation performance across candidates. The reward encourages task completion and smooth motion while penalizing undesired contacts and object toppling.

Stage 3: Critic Distillation for Task-Aware Grasp Selection

The PPO critic estimates the expected discounted return from its critic observation, implicitly encoding the downstream utility of each grasp: configurations that enable more successful, efficient, and stable rearrangement receive higher value estimates. However, the critic uses privileged simulator information, including physical properties such as object mass and velocity, that is unavailable before interaction at deployment.

Grasp-TAGS distills these grasp-dependent critic values into a lightweight utility network that uses only the candidate grasp configuration and target object pose in the initial object frame. After policy training, the network regresses its utility predictions to critic values from sampled initial states. At inference time, it evaluates every candidate, selects the grasp with the highest predicted utility, and executes the learned manipulation policy conditioned on that grasp.

Summary

Grasp-TAGS combines a stable hierarchical loco-manipulation policy, grasp-conditioned PPO training, and critic distillation to turn long-horizon task value into a practical pre-interaction grasp decision. This lets the robot choose the grasp expected to best support the desired large-object rearrangement before policy execution.