DiPLO: Self-distilled Diffusion Planning for Loco-manipulation
Abstract Overview
DiPLO is a humanoid loco-manipulation framework that combines a diffusion-based motion planner with a low-level reinforcement learning tracking policy. The planner is pretrained on dynamically feasible robot-object motions derived from a small set of retargeted demonstrations, and then improved in two post-training stages. First, SD:DPO self-distills a reference-conditioned planner into one that operates using state history and object-goal conditioning alone. Second, DA:PPO fine-tunes the distilled planner through closed-loop interaction with a frozen tracker to improve task completion and align planned motions with executable behavior.
Novelty
The paper introduces a two-stage post-training pipeline combining self-distillation (SD:DPO) and diffusion-RL fine-tuning (DA:PPO) to eliminate deployment reliance on kinematic motion priors while preserving their benefits during pretraining. It also models rotational components directly on SO(3) within the diffusion planner and demonstrates a unified planner-tracker architecture across pick-place, kicking, pushing/pulling, and real-robot hardware transfer.
Results
In pick-carry-place simulation, SD:DPO raises the goal-conditioned student's survival rate from 57.7% to 100.0%, and subsequent DA:PPO fine-tuning achieves 99.5% survival and 98.8% task success. For the augmented pretraining setup, DA:PPO increases task success from 51.9% to 96.4%, outperforming both the pretrained model and a DPPO-style baseline in tracking accuracy and object-pose errors. The distilled system also achieves 88% task success across a polar grid of unseen goals and transfers zero-shot to a Unitree G1 robot with 10/10 repeated and 5/5 distinct successful executions.
Key Points
- Strong inverse-kinematics task conditioning is essential during pretraining in limited-data regimes, but SD:DPO successfully removes this requirement at deployment by distilling the reference-conditioned teacher into a goal-conditioned student.
- DA:PPO fine-tuning substantially improves closed-loop task completion over the pretrained model and outperforms DPPO-style updates, while planner fine-tuning achieves higher task success than fine-tuning the tracking policy alone.
- A hierarchical motion-planning-plus-tracking architecture outperforms direct action-level diffusion baselines in survival and task completion, generalizing robustly to out-of-distribution goals and real humanoid hardware.