PaperlayerView exact revision on arXiv

Source-linked AI summary

Collision-Free Humanoid Traversal in Cluttered Indoor Scenes

Han Xue, Sikai Liang, Zhikai Zhang, Zicheng Zeng, Yun Liu, Yunrui Lian, Jilong Wang, Qingtao Liu, Xuesong Shi, Li Yi

arXiv:2601.16035v2cs.RO

TL;DR

Collision-free traversal in cluttered indoor scenes requires humanoids to map varied obstacle layouts and geometries to traversal skills, but existing representations do not effectively encode humanoid–obstacle relationships. The paper proposes HumanoidPF and hybrid scene generation for learning those skills. Experiments in simulation and realistic real-world indoor scenes validate practical applicability and generalization across diverse environments.

  • Problem

    Humanoids must map varied obstacle layouts and geometries to traversal skills, but existing approaches lack an effective representation of humanoid–obstacle relationships during collision avoidance.

    Source: §I · pp. 1–2

  • Method

    HumanoidPF encodes collision-free motion directions for RL-based traversal learning, while hybrid scene generation combines realistic indoor-scene crops with procedurally synthesized obstacles.

    Source: §I · p. 2

  • Results

    Experiments in simulation and realistic real-world indoor scenes validate HumanoidPF's practical applicability and strong generalization across diverse environments.

    Source: §I · p. 2

  • Takeaways & Limitations

    The approach advances collision-free humanoid traversal toward real-world domestic robot applications and includes a system for single-click commands in cluttered indoor environments.

    Source: §V · p. 7

    The framework does not exploit contact-rich interactions, and generalization to entirely unseen, highly unstructured, and extremely cluttered scenes remains an open challenge.

    Source: §V · p. 8

Abstract

from arXiv · show

We study the problem of collision-free humanoid traversal in cluttered indoor scenes, such as hurdling over objects scattered on the floor, crouching under low-hanging obstacles, or squeezing through narrow passages. To achieve this goal, the humanoid needs to map its perception of surrounding obstacles with diverse spatial layouts and geometries to the corresponding traversal skills. However, the lack of an effective representation that captures humanoid-obstacle relationships during collision avoidance makes directly learning such mappings difficult. We therefore propose Humanoid Potential Field (HumanoidPF), which encodes these relationships as collision-free motion directions, significantly facilitating RL-based traversal skill learning. We also find that HumanoidPF exhibits a surprisingly negligible sim-to-real gap as a perceptual representation. To further enable generalizable traversal skills through diverse and challenging cluttered indoor scenes, we further propose a hybrid scene generation method, incorporating crops of realistic 3D indoor scenes and procedurally synthesized obstacles. We successfully transfer our policy to the real world and develop a teleoperation system where users could command the humanoid to traverse in cluttered indoor scenes with just a single click. Extensive experiments are conducted in both simulation and the real world to validate the effectiveness of our method. Demos and code can be found in our website: https://axian12138.github.io/CAT/.

I. INTRODUCTION

Humanoids in cluttered indoor scenes must map varied obstacle layouts and geometries to traversal skills while avoiding collisions. The paper introduces HumanoidPF and hybrid scene generation, and reports validation in simulation and real-world scenes.

Motivation

  • The paper identifies a gap in collision avoidance: existing methods often provide only sparse collision penalties or raw environmental measurements, leaving humanoid–obstacle relationships implicit.These limitations make traversal decisions difficult to learn for scenes with full-spatial layouts and intricate geometries.Source: §I · pp. 1–2

HumanoidPF

  • HumanoidPF represents humanoid–obstacle relationships as a continuous field whose virtual forces indicate collision-free motion directions.It is queried at key body parts for policy observations, and its preferred motion directions guide collision-aware rewards for RL.Source: §I · p. 2

Training environments

  • Hybrid scene generation combines crops of realistic 3D indoor datasets with procedurally synthesized difficult obstacles to expose policies to challenging clutter configurations.The method targets near-collision and emergency scenarios that are rarely present in existing datasets.Source: §I · p. 2

Deployment and validation

  • Click-and-Traverse lets a user command a humanoid to traverse cluttered indoor environments by clicking a goal.The authors report that simulation and realistic real-world indoor experiments validate HumanoidPF's practical applicability and generalization across diverse environments.Source: §I · p. 2

Contribution

  • The paper presents collision-free humanoid traversal in cluttered indoor scenes as a step toward domestic robot applications.It describes this as the first systematic study of the problem, to the authors’ knowledge.Source: §I · pp. 1–2

II. RELATED WORKS

Prior legged-robot work includes challenging locomotion, but humanoid studies often address only partial obstacle layouts or simple geometries. The paper positions HumanoidPF as a reformulation of artificial potential fields for whole-body humanoid traversal.

A. Legged locomotion in complex environments

  • Humanoid studies have addressed height-constrained settings and risky terrains, but often use partial spatial layouts or simple obstacle geometries.Gallant [29] considers ground, lateral, and overhead layouts separately rather than with all constraints coexisting.Source: A. Legged locomotion in complex… · p. 2

B. Artificial potential field for obstacle avoidance

  • Artificial potential fields guide obstacle avoidance with virtual forces, but prior quadruped approaches abstracted the robot as a single rigid body, such as its center of mass or foot joint.HumanoidPF reformulates APF for humanoid skill learning in cluttered scenes with full-spatial constraints and intricate geometries.Source: B. Artificial potential field for… · pp. 2–3, Figure 2 · p. 3

III. METHOD

The method addresses reaching a target without collisions by mapping obstacle perception to traversal skills. It presents HumanoidPF for whole-body traversal learning and situates it alongside scene generalization and teleoperated deployment.

III. METHOD

  • The method also targets generalization to diverse, challenging indoor scenes and instantiates real-world deployment as teleoperated loco-navigation.The overall method is divided between HumanoidPF learning, scene generalization, and teleoperated deployment.Source: §III · p. 3

A. HumanoidPF for whole-body traversal learning

  • HumanoidPF reformulates artificial potential fields for whole-body traversal, encoding humanoid–obstacle relationships for informative perception and reward streamlining.The method targets movement to a goal without colliding with indoor obstacles.Source: A. HumanoidPF for whole-body traversal… · p. 3, §III · p. 3

1) HumanoidPF construction

  • The combined potential’s negative gradient forms a guidance field that is queried at body-part locations, using obstacle-aware geodesic distance and collision-preventing repulsion.Geodesic distance represents the shortest 3D path to the goal without intersecting obstacles.Source: 1) HumanoidPF construction · pp. 3–4
  • Symmetric obstacle fields can cancel across body parts, trapping the robot in local minima or causing oscillation; priority weighting adjusts body-part influence to address these conflicts.The root receives higher priority for stability and direction, while collision urgency depends on signed distance and body-part velocity.Source: 1) HumanoidPF construction · p. 4

2) Traversal skill learning with HumanoidPF

HumanoidPF gives the policy body-part-level directional cues for collision-free traversal and supplies a preferred-motion distribution for reward design. Its reward formulation generalizes across scenes without manual tuning, while its observation representation mitigates perceptual sim-to-real differences.

Traversal skill learning with HumanoidPF

  • Sampling HumanoidPF at 13 body parts gives the policy local collision-free steering guidance for whole-body traversal.The field combines obstacle and target guidance, letting the policy reason about traversal decisions rather than infer them from raw visual data.Source: 2) Traversal skill learning with… · p. 4
  • The continuous, spatially aggregated HumanoidPF observation suppresses isolated noise while retaining traversal-relevant gradients, mitigating perceptual sim-to-real differences.The representation is intended to make minor geometric variations less influential during real-world control.Source: 2) Traversal skill learning with… · p. 4
  • HumanoidPF models preferred motion directions with a von Mises–Fisher distribution, with concentration controlling the strength of directional alignment.Higher-priority body parts receive greater field magnitude and concentration, promoting coordinated whole-body motion and stricter alignment.Source: 2) Traversal skill learning with… · p. 4
  • The HumanoidPF reward generalizes across scenes without manual tuning, supporting an automated training pipeline that scales across diverse environments.Source: 2) Traversal skill learning with… · p. 5

B. Scalable training in diverse and challenging scenes

A unified policy must generalize across diverse, challenging indoor scenes, but directly training one policy across all scenes remains difficult because of RL sample inefficiency. The proposed approach pairs hybrid scene generation with specialist-to-generalist training.

B. Scalable training in diverse and challenging scenes

1) Hybrid scene generation

The hybrid scene-generation method combines realistic 3D indoor crops with procedurally generated obstacles to address the scarcity of highly challenging layouts. A difficulty-controlled curriculum expands obstacle complexity during training.

1) Hybrid scene generation

  • Highly challenging obstacle layouts are a long-tail subset of existing datasets, so simply scaling dataset size does not address their scarcity.Typical indoor scenes have orderly object arrangements and clearly delineated walkable regions.Source: 1) Hybrid scene generation · p. 5
  • Crops from realistic 3D-FRONT scenes provide structurally realistic indoor environments for policy training.Scene blocks are selectively cropped and filtered.Source: 1) Hybrid scene generation · p. 5
  • Procedurally generated boxes create highly restrictive layouts with simultaneous ground, lateral, and overhead constraints.Boxes vary in position, dimensions, and orientation, and can form narrow passages.Source: 1) Hybrid scene generation · p. 5
  • A layout-agnostic difficulty factor progressively increases obstacle complexity, such as box number and size, during curriculum learning.Source: 1) Hybrid scene generation · p. 5

2) Specialist-to-generalist training

The training pipeline first learns specialist policies across diverse scenes, then distills a generalist from those specialists. Training also uses sensor noise, force perturbations, and progressively harder scenes.

2) Specialist-to-generalist training

  • Specialist policies are trained with PPO across 139 cropped 3D-FRONT scenes and 216 procedurally generated scenes using 32,768 parallel environments and 5,000,000 episodes each.Start and goal locations are randomly sampled for each episode.Source: 2) Specialist-to-generalist training · p. 6
  • DAgger distills a generalist from multiple specialist teachers, whose expert actions are conditioned on the current obstacle.Source: 2) Specialist-to-generalist training · p. 6
  • Both specialist and generalist policies train with sensor noise, force perturbations, and a curriculum of increasing scene difficulty.The curriculum is used to enhance traversal performance and explore policy limits.Source: 2) Specialist-to-generalist training · p. 6

C. Real-world deployment: Click-and-Traverse

Click-and-Traverse lets users click a goal on a grid map, then autonomously navigate while avoiding collisions. The experiments evaluate traversal in simulation and on a Unitree G1 humanoid robot.

C. Real-world deployment: Click-and-Traverse

  • The real-world system uses LiDAR–inertial SLAM and OctoMap to update its environment map and HumanoidPF construction at 10 Hz.The policy queries HumanoidPF at different body parts using self-localization and forward kinematics.Source: C. Real-world deployment:… · p. 6
  • Users click a goal on a grid map, and the humanoid autonomously navigates to it while avoiding collisions.The interface avoids labor-intensive control modalities such as joysticks or motion capture.Source: C. Real-world deployment:… · p. 6

IV. EXPERIMENT

  • Experiments use MuJoCo and the Unitree G1 to test HumanoidPF against existing methods, hybrid-scene generalization, and sim-to-real transfer.These are the three questions the experiments are designed to address.Source: §IV · p. 6

A. Validation of HumanoidPF for skill learning

Across the tested cluttered scenes, the method achieves the best performance among the compared methods, with low variance across trials. The evaluation spans eight manually generated scene types and includes comparisons of perception and reward choices.

A. Validation of HumanoidPF for skill learning

  • Evaluation uses eight manually generated cluttered-scene types, with 10 scenes per type, to cover varied obstacle layouts and geometries.All methods are trained and evaluated on these generated scenes.Source: A. Validation of HumanoidPF for skill… · p. 6
  • Success requires reaching within 0.1 meters of the target in 5 seconds without collision, while Distance Error measures the average closest horizontal distance from the humanoid root to the target.Source: A. Validation of HumanoidPF for skill… · p. 6
  • Comparisons include ASTraversal and Humanoid Parkour, plus ablations replacing HumanoidPF reward with collision penalties or its observation with multi-layer elevation maps.The baselines are adapted for the humanoid setting, including overhead obstacle perception for Humanoid Parkour.Source: A. Validation of HumanoidPF for skill… · p. 7
  • The method achieves the best performance across all tested case types, while Humanoid Parkour and ASTraversal are competitive mainly on simpler terrains.The authors also report low variance over multiple trials, indicating robustness and stability.Source: A. Validation of HumanoidPF for skill… · p. 7

B. Validation of hybrid scene generation

Procedurally generated obstacles improve generalization to unseen scenes, with the strongest reported gains associated with training on novel, highly constrained layouts. The evaluation compares policies on artist-designed scenes excluded from training.

B. Validation of hybrid scene generation

  • The zero-shot evaluation uses 30 artist-designed unseen scenes, split into 15 easy and 15 hard cases.Policies trained with different scene datasets are compared on these scenes.Source: B. Validation of hybrid scene generation · p. 7
  • Training variants combine 3D-FRONT scenes with subsets or the full set of synthesized scenes, including a high-difficulty set scored from 0.4 to 1.0.The compared variants include moderate-difficulty synthesized scenes as well as the high-difficulty set.Source: B. Validation of hybrid scene generation · p. 7
  • Procedural obstacle generation improves generalization, with the key breakthrough emerging from high-difficulty scenes that introduce novel complexities and tighter constraints.Gains can saturate once common layout patterns are sufficiently learned.Source: B. Validation of hybrid scene generation · p. 7

C. Validation of HumanoidPF for real world transfer

In real-world cluttered scenes, HumanoidPF consistently outperforms voxel-based perception, and the authors report behavior closer to simulation despite disturbances. The conclusion notes remaining challenges in contact-rich interactions and extremely cluttered, unstructured unseen scenes.

C. Validation of HumanoidPF for real world transfer

  • The real-world tests stress perception during crouching under low obstacles, hurdling over floor objects, and negotiating narrow-side passages.Source: C. Validation of HumanoidPF for real… · p. 7
  • HumanoidPF consistently outperforms voxel-based approaches across diverse real-world environments, with more stable hurdle behavior under geometric detail and task-irrelevant noise.The authors report that its behavior aligns more closely with simulation.Source: C. Validation of HumanoidPF for real… · p. 7

V. CONCLUSION

  • The framework does not yet exploit contact-rich actions such as leaning on obstacles or stepping onto support surfaces, and generalization to entirely unseen, highly unstructured, extremely cluttered scenes remains open.Source: §V · pp. 7–8
Loading 2601.16035v2…