HIPC

Humanoid Iterative Planning and Control
via Goal-Conditioned Motion Closure

HIPC: Humanoid Iterative Planning and Control
via Goal-Conditioned Motion Closure

Yi Han1*, Xingyu Chen2,3*, Zixuan Lang3, Qianyang Wu4, Yuhang Lin5, Ji Ma6, Liang Li7,
Dongting Li7, Xinpeng Liu1, Ruoli Dai3, Hesheng Wang1†, Renjing Xu2†, Lei Han3†
*Equal Contribution    †Equal Advising
1 SJTU    2 HKUST(GZ)    3 Noitom Robotics    4 ShanghaiTech    5 ZJU    6 HKU    7 THU
PDF PDF arXiv arXiv GitHub Code (coming soon)
HIPC teaser

Abstract

Recent generative models are widely used to provide whole-body motion for humanoid tracking. Current methods combine motion generators with controllers in an open-loop paradigm, which leads to critical issues: tracking errors can accumulate into spatial drift, and the planner cannot adapt to unexpected physical disturbances without feedback from physical execution. However, directly feeding realized states back to the planner in a closed loop can cause discontinuities, resulting in physical jitter and even falls, since such states are out of distribution during training. To bridge this gap, we propose HIPC, a real-time Humanoid Iterative Planning and Control framework that couples a generative model with motion tracking into a unified planning-and-control paradigm. HIPC introduces a goal-conditioned motion planning trained with a canonical motion representation that avoids gimbal-lock singularities under extreme orientations and supports motion generation on diverse geometric and semantic conditions. Then, HIPC adopts two strategies, deviation-decayed recovery and history re-anchoring, to mitigate the out-of-distribution shift caused by realized-state feedback. Finally, HIPC operates a real-time asynchronous planning-and-control system, with latency compensation to reduce temporal mismatch for inference delay. Extensive experiments confirm the effectiveness of HIPC under switching and composition of diverse conditions. Results show that HIPC achieves state-of-the-art goal-reaching accuracy among existing methods while simultaneously demonstrating robustness to physical perturbations and the ability to recover from falls.

1. Goal-Conditioned Tracking

HIPC generates whole-body motion references from geometric and semantic conditions, and closes the loop with physical execution. These examples show goal reaching, online goal updates, and real-world deployment.

1.1 Individual Goals

HIPC supports geometric goal conditions at different levels of the humanoid state and body configuration.

#1 Root Position

HIPC reaches root-position goals across different target distances and heights.

Goal Distance: 1 m
Goal Distance: 2 m
Goal Distance: 3 m
Goal Height: 0.4 m
Goal Height: 0.6 m
Goal Height: 0.8 m

#2 Root Orientation

Yaw: 90°
Yaw: 180°
Pitch: 45°
Pitch: -45°
Roll: 30°
Roll: -30°
Roll: 30°, Pitch: 40°
Pitch: 45°, Yaw: -90°

#3 Root Velocity

Forward: 0.6 m/s
Forward: 1.0 m/s
Forward: 2.5 m/s
Backward: 1.0 m/s
Left: 1.0 m/s
Right: 1.0 m/s

#4 Joint Configuration

Half Squat
Upper Body Stretch

#5 End-Effector Position

Static Goal
Dynamic Goal

1.2 Composed Goals

Running
Marching

Text as a Motion Prior. Text conditions guide the style and behavior of generated whole-body motion. These examples show running and marching priors used alongside geometric goal conditions.

1.3 Static and Switching Goal Reaching

Goal Attraction
Dynamic Goal Switching

Static and switching goal reaching. Left: Trajectories from diverse initial states converge to a shared fixed goal. Right: The robot continuously adapts its motion to switching goals.

1.4 Real-World Goal Tracking

Open-Loop
Closed-Loop (HIPC)

2. Robust Tracking & Recovery

Execution feedback allows HIPC to revise motion plans when physical motion deviates from the reference. The following examples distinguish trajectory correction from recovery after a fall.

2.1 Robustness under Perturbations

HIPC corrects spatial drift and maintains the desired motion under persistent physical perturbations.

2.2 Fall Recovery & Get-Up

The humanoid recovers from falls caused by large disturbances and resumes tracking the desired motion.

Real-World Recovery. The humanoid recovers from a fall and returns to a standing posture.

3. Comparison Experiments

3.1 Baseline Comparison

We compare HIPC with TextOp, OMG, Kimodo, and ARDY. For each comparison case, both the input motion history and the future root positions used as planar goals are sampled from 100STYLE. All methods are connected to the same SONIC tracker under the same closed-loop protocol. The videos below show the resulting planning-and-control executions after SONIC tracking, alongside the corresponding ground-truth motions.

3.2 Open-Loop vs. Closed-Loop

In the open-loop setting, the planner autoregressively generates motion based on its own predicted history, without feedback from physical execution. This causes spatial drift to accumulate, leading to an increasing error between the realized state and the reference motion. HIPC instead feeds back the realized state to re-anchor the planning history and update future motion references, enabling correction when physical execution deviates from the plan.

Open-Loop
Planning from predicted history
Closed-Loop (HIPC)
Replanning with realized state

The solid G1 shows the motion under physics simulation, the semi-transparent G1 shows the reference motion generated online by the planner, and the blue G1 indicates the goal.

4. Scene Navigation

We demonstrate HIPC with two upstream navigation modules: an occupancy-map-based A* planner with obstacle inflation for maze navigation, and the vision-and-language navigation model OmniNav for indoor navigation. Their waypoints are transformed into HIPC's egocentric goal conditions for online navigation.

4.1 A*-Guided Maze Navigation

Starting near the center of the maze, the humanoid follows waypoints from an occupancy-map-based A* planner with obstacle inflation through the corridors toward the exit.

4.2 VLN-Guided Indoor Navigation

Overview of VLN-guided indoor navigation

We evaluate VLN-guided indoor navigation in six Replica scenes imported into Isaac Sim. OmniNav uses visual observations and language instructions to provide dynamically updated waypoints, which HIPC executes as egocentric goals. Each example below pairs an input frame with the humanoid execution video.

Scene experiment 1

Case #1 — Instruction: Go straight, step out of the door, and stop immediately.



Scene experiment 2

Case #2 — Instruction: Walk to the right and forward, then enter the first door on the right and stop.



Scene experiment 3

Case #3 — Instruction: Go straight and walk into the open door on the right.



Scene experiment 4

Case #4 — Instruction: Go straight, passing through the narrow gap between the blue chair and the blue low sofa. Then stop.



Scene experiment 5

Case #5 — Instruction: Go straight, passing the bed, and stop.



Scene experiment 6

Case #6 — Instruction: Go straight and stop in front of the dark blue sofa.

Conclusion

We presented HIPC, a closed-loop humanoid planning-and-control framework that integrates goal-conditioned generative motion planning with execution-consistent feedback and latency-aware asynchronous control. By closing the loop between motion generation and physical execution while preserving planner–controller modularity, HIPC supports accurate goal reaching, robust perturbation recovery, and recovery from unseen fallen configurations across diverse whole-body tasks.

BibTeX

@article{han2026hipc,
  title={HIPC: Humanoid Iterative Planning and Control via Goal-Conditioned Motion Closure},
  author={Han, Yi and Chen, Xingyu and Lang, Zixuan and Wu, Qianyang and Lin, Yuhang and Ma, Ji and Li, Liang and Li, Dongting and Liu, Xinpeng and Dai, Ruoli and Wang, Hesheng and Xu, Renjing and Han, Lei },
  journal={arXiv preprint arXiv:2610.41334},
  year={2026}
}