DiPLO

Self-distilled Diffusion Planning for Loco-manipulation

A diffusion planner generates whole-body robot and object motion for a future horizon, and a single RL tracking policy executes it. Two post-training stages ensue: Self-Distillation Diffusion Policy Optimization (SD:DPO) removes the need for strong task conditioning that is required for strong pretraining performance, allowing us to plan for motions using the robot's state history and goal alone, and Diffusion-Augmented PPO (DA:PPO) allows us to finetune the diffusion planner to achieve maximum task success. The resulting planner-tracker pipeline is deployed on the real robot without any real-world finetuning.

  • Shafeef Omar1,4
  • Sebastian Sanokowski1
  • Elham Daneshmand2
  • Dian Yu1
  • Daniel Cremers1
  • Cordelia Schmid3,4
  • Majid Khadiv1
1Technical University of Munich 2McGill University 3Inria 4Institute for Advanced Study, TUM
Paper arXiv Code Watch the film BibTeX
The same motion in simulation (left, led by the cyan hologram of the planner's plan) and on the Unitree G1 (right).
2.7 min

Of retargeted human motion used from the existing OmniRetarget dataset

45% → 99%

Task success in pick-carry-place using state-history and goal conditioning only, from the pretrained planner to SD:DPO + DA:PPO

The film · 3:36

Method, simulation results and the real robot deployment. Watch on YouTube.

Method

The robot's path seen from above in every training motion, all starting at the origin facing right. Hover or tap a path to pick out every version of its motion. Quarter turns of the object leave the robot's path unchanged, so the symmetry view adds the mirrored paths.

From 29 human motions to 7,184 feasible robot motions

Each human motion is retargeted to the robot by inverse kinematics (IK) based methods. This retargeting ignores dynamics and contact forces. Sampling-based trajectory optimization (SBTO) re-solves every clip in simulation from perturbed initial conditions and gives several dynamically feasible solutions for it. Mirroring and quarter turns of the object multiply the set again.

  • 90.0 → 96.5%pretrained survival with one motion per clip vs. SBTO variations and symmetry
  • 72.3 → 96.4%task success after DA:PPO with one motion per clip vs. SBTO variations and symmetry

Why do we use IK-retargeted robot motion as task conditioning?

Survival of the pretrained planner under different task conditioning.% of rollouts with the frozen tracker; whiskers ± SEM over motions
plan 1 / 9Drag to orbit
0.0 s
Rollouts of the final planner for different goals. The cyan holograms are the plan being executed, 0.32 to 0.96 s ahead of the robot; the ticks on the timeline mark each new plan.

Generalization

Unseen goals. One start, goals on a 2 m ring around it, none from the human motions.
Goals moved 0.5 m and 1 m forward, back, left and right of where the human placed the object.
Goals turned in place. The robot walks around to set the object down turned by 90°.

The goal moves during the task and the plan follows it: slides, turns on the spot, arcs, a zig-zag, and backwards with a 180° turn.

Goal 20 m away: Ten times farther than any goal from our human motions dataset.

Other skills

Student + SD:DPO + DA:PPO, conditioned on the state history and the goal only. The cyan brackets mark the goal pose.

Real world

Executions on the real robot from different start poses and toward different goals.

Cite

BibTeX

@misc{omar2026diplo,
  title         = {{DiPLO}: Self-distilled Diffusion Planning for Loco-manipulation},
  author        = {Omar, Shafeef and Sanokowski, Sebastian and Daneshmand, Elham and
                   Yu, Dian and Cremers, Daniel and Schmid, Cordelia and Khadiv, Majid},
  year          = {2026},
  eprint        = {2610.24847},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.24847}
}