PaperlayerView exact revision on arXiv

Source-linked AI summary

HOTICE: Whole-Body Humanoid Object Transportation in Cluttered Environments

Toan Nguyen, Weiduo Yuan, Siheng Zhao, Yue Wang, Daniel Seita

arXiv:2609.25363v1cs.RO

TL;DR

Cluttered object transport challenges methods that do not account for both the humanoid and its payload, while whole-body control creates a large action space. HOTICE combines decoupled guidance fields, dual-agent control, and teacher-to-student distillation; it reports strong performance in simulation, real-world rollouts, and previously unseen scenes. Its scope excludes moving obstacles and post-collision recovery modeling.

  • Problem

    Clutter constrains free space around both the humanoid and its payload, while coordinating whole-body motion and maintaining a secure hold makes transport challenging.

    Source: §I · p. 1

  • Method

    HOTICE combines humanoid-object decoupled guidance fields, dual-agent upper- and lower-body control, and distillation of privileged scene-specific teachers into a generalist policy.

    Source: §I · p. 2

  • Results

    HOTICE outperforms other baselines on both reported metrics and, in unseen scenes, averages 80.1% success rate and 0.27 m mDist while demonstrating real-world performance.

    Source: A. Simulation Experiments · p. 6, B. Real-World Results · p. 7

  • Takeaways & Limitations

    HOTICE demonstrates reliable, collision-aware transport across diverse clutter and object shapes, including previously unseen scenes and sim2real transfer.

    Source: §VII · p. 8

    The method does not model post-collision scenarios or dynamic scenes with moving obstacles, and complex carrying strategies remain out of reach.

    Source: §VI · p. 8

Abstract

from arXiv · show

Object transportation is a fundamental capability for humanoid robots operating in real-world, human-centric environments, yet existing methods struggle when clutter constrains free space around both the robot and its carried payload. We present HOTICE, a whole-body humanoid learning framework for transporting objects through such cluttered environments. First, we introduce Humanoid-Object Decoupled Potential Fields, which jointly encode collision-avoidance guidance for the robot and the carried object, enabling coordinated, obstacle-aware motion for both. Second, to address the large action space inherent to whole-body loco-manipulation, we design a dual-agent reinforcement learning architecture that decouples upper- and lower-body control while preserving whole-body coordination via shared state observations and rewards. To train a policy that generalizes across diverse cluttered scenes, we further employ a specialist-to-generalist distillation strategy, in which privileged teacher policies are distilled into a single deployable student policy. We evaluate HOTICE in MuJoCo simulation and on a real Unitree G1 humanoid, demonstrating effective and robust object transportation across cluttered scenarios for objects of varying shapes. Our results show that HOTICE reliably coordinates whole-body motion and object-aware collision avoidance, generalizing effectively to previously unseen cluttered environments while achieving strong performance in sim2real deployment.

I. INTRODUCTION

Traditional robot platforms rely on assumptions that fail in clutter, while humanoid object transport requires coordinated whole-body control and secure payload handling. HOTICE combines object-aware guidance, dual-agent control, and policy distillation for transport across cluttered scenarios.

I. INTRODUCTION

  • Fixed-base arms cannot reposition around obstacles, wheeled bases struggle with thresholds and narrow passages, and drones have limited payload capacity and need open airspace.These platforms achieve high reliability in structured, engineered settings.Source: §I · p. 1
  • Cluttered humanoid transport requires coordinated upper- and lower-body control, balance in extreme poses, and a secure hold on objects that may be heavy or large.The combined demands create a large, high-dimensional action space and reduce clearance.Source: §I · p. 1
  • HOTICE uses Humanoid-Object Decoupled Potential Fields to guide joint avoidance of humanoid-obstacle and object-obstacle collisions.The fields maintain synchronization between the robot and transported object.Source: §I · p. 2
  • A dual-agent learning architecture decouples upper- and lower-body action spaces to address the large action space of whole-body control.The architecture is presented as improving loco-manipulation capability in complex environments.Source: §I · p. 2
  • HOTICE distills scene-specific privileged teacher policies into a generalist student and reports transport across scenarios in simulation and real-world deployment.The system is evaluated on a Unitree G1 with objects of different shapes in MuJoCo and real-world settings.Source: §I · p. 2

II. RELATED WORK

Prior humanoid traversal methods address the robot’s own collision geometry, but object transport extends that geometry and perturbs balance. HOD-PF addresses this gap with separate, jointly acting guidance fields for the humanoid and payload.

A. Humanoid Traversal in Cluttered Environments

  • Legged traversal research includes model-based planning and learned policies, while humanoid work includes parkour, contact-driven control, balance, and terrain-conditioned skill composition.Humanoid traversal is considered harder because of higher dimensionality and narrower stability margins.Source: A. Humanoid Traversal in Cluttered… · p. 2
  • HumanoidPF and TANGO address cluttered traversal but assume an unencumbered robot whose collision geometry follows its own kinematics.Transported payloads extend the collision volume and perturb balance when extreme postures may be required.Source: A. Humanoid Traversal in Cluttered… · p. 2
  • HOD-PF decouples humanoid and object collision avoidance into separate guidance vector fields that jointly shape carrying motion through clutter.This targets the gap left by robot-only traversal methods.Source: A. Humanoid Traversal in Cluttered… · p. 2

B. Humanoid Object Transportation

Prior object-transport methods largely target open or lightly obstructed settings, while HOTICE addresses clutter that constrains both robot and payload. The task requires collision-free transport while holding an object, and the method combines guidance fields, dual-agent control, and teacher-to-student distillation.

B. Humanoid Object Transportation

  • Prior planning and learning-based humanoid transport methods address open or lightly obstructed scenes, whereas real clutter constrains the robot and payload from the ground, sides, and above.HOTICE targets transportation through complex, highly constrained scenes.Source: B. Humanoid Object Transportation · p. 2

III. PROBLEM STATEMENT

  • A successful task reaches the goal while the robot still holds the object, with no robot-object collision; falls, drops, and collisions count as failures.The object is assumed already held for the main task description and experiments, though a pickup policy is also trained.Source: §III · p. 2

IV. METHOD

  • HOTICE combines obstacle-aware guidance for robot and object, decoupled upper- and lower-body control, and distillation of scene-specific teachers into one generalizable policy.The overview specifies unified state observations and shared whole-body rewards for the body-specific actors, followed by DAgger and RL fine-tuning for the student.Source: §IV · pp. 2–3, Figure 2 · p. 3

A. Humanoid-Object Decoupled Potential Fields

HOD-PF combines robot-body guidance with object-centric guidance to coordinate collision avoidance during carrying. The object field also accounts for the payload’s clearance and requires obstacle-driven hand and object adjustments beyond ordinary locomotion.

A. Humanoid-Object Decoupled Potential Fields

  • HOD-PF extends robot-only HumanoidPF with separate guidance fields for robot traversal and object motion, synchronizing the object with the robot.The Standard Guidance Field guides the head, pelvis, and feet; the Object-Centric Guidance Field guides the object and hands.Source: A. Humanoid-Object Decoupled Potential… · p. 3
  • The Standard Guidance Field combines an attractive goal potential and a repulsive obstacle potential, then supplies guidance vectors for robot body parts.Its negative gradient points toward the goal while pushing away from obstacles; rewards encourage head, pelvis, and feet velocities to align with their guidance vectors.Source: A. Humanoid-Object Decoupled Potential… · pp. 3–4
  • The Object-Centric Guidance Field computes guidance relative to obstacles inflated by a margin, adding robustness to sensor noise and control delay and encouraging earlier object repositioning.The margin balances collision avoidance and traversal feasibility.Source: A. Humanoid-Object Decoupled Potential… · p. 4
  • Object keypoint and hand guidance vectors enter the policy, while the steering reward compares their residual motion after subtracting the overall locomotion command.This credits obstacle-driven adjustments assigned to the hands rather than motion already produced by locomotion.Source: A. Humanoid-Object Decoupled Potential… · p. 4

B. Dual-Agent Learning Architecture

HOTICE uses a dual-agent architecture to address the difficulty of controlling a humanoid and carried object through complex scenes. The agents split upper- and lower-body control but remain coupled through shared observations, rewards, and joint optimization.

B. Dual-Agent Learning Architecture

  • The Object-Centric Guidance Field provides earlier collision-avoidance signals and more headroom than the Standard Guidance Field.These specialized signals address the challenge of traversing complex geometries while controlling an unwieldy object.Source: Figure 3 · p. 4, B. Dual-Agent Learning Architecture · p. 4
  • Separate agents control the 12 leg joints and 8 arm joints, emitting actions that are concatenated into a 20D action for the PD controller.The agents read the same observation and share termination conditions; waist and wrist joints are excluded.Source: B. Dual-Agent Learning Architecture · p. 4
  • Each agent forms its own advantage estimate and PPO objective, while shared reward terms keep both accountable for carrying.Locomotion and balance rewards go to the lower-body agent; reaching and grasp-maintenance rewards go to the upper-body agent.Source: B. Dual-Agent Learning Architecture · p. 4

C. Specialist-to-Generalist Training

HOTICE trains scene-specific teachers and distills them into a deployable generalist student, then refines the student with PPO. Training spans procedurally generated and realistic indoor scenes without scene-specific reward tuning.

C. Specialist-to-Generalist Training

  • Parallel teachers train on scene-specific privileged states, using both procedurally generated scenes and realistic indoor scenes cropped from 3D-FRONT.The generated scenes push policy limits, while realistic scenes support generalization to real-world scenarios.Source: C. Specialist-to-Generalist Training · p. 4
  • DAgger distillation combines the teacher policies into one deployable generalist student, which is further refined with PPO and simulated sensor noise and control delays.The student follows the same dual-agent framework as its teachers.Source: C. Specialist-to-Generalist Training · p. 5
  • Training uses no scene-specific reward tuning, and qualitative results include procedurally generated and realistic indoor scenes.The figure caption also reports balance and stability under challenging obstacles and more reliable collision avoidance in its comparisons.Source: C. Specialist-to-Generalist Training · p. 5, Figure 4 · p. 5

V. EXPERIMENTS

The authors conduct experiments in simulation and in the real world to validate HOTICE’s performance and generalizability.

V. EXPERIMENTS

  • Simulation experiments are part of the evaluation of HOTICE’s performance and generalizability.The paper also evaluates the method in the real world.Source: §V · p. 5
  • Real-world experiments are part of the evaluation of HOTICE’s performance and generalizability.The paper also evaluates the method in simulation.Source: §V · p. 5
  • The experiments are intended to validate both performance and generalizability.The evaluation spans simulation and the real world.Source: §V · p. 5

A. Simulation Experiments

In simulation, HOTICE outperforms its ablations on both reported metrics, maintains strong performance on unseen scenes, and performs consistently across box, cylinder, and sphere objects.

A. Simulation Experiments

  • The ablations remove the Object-Centric Guidance Field, the dual-agent architecture, or both to assess the contributions of HOTICE’s components.The field-only baseline applies the Standard Guidance Field to keypoints on both the robot and carried object.Source: A. Simulation Experiments · p. 5
  • Each method trains teachers for 400M timesteps on 20 diverse scenes, then distills them into a student policy for evaluation.Evaluation applies input noise, delays, and perturbations and benchmarks 20 scenes with 1,000 rollouts each.Source: A. Simulation Experiments · pp. 5–6
  • HOTICE outperforms the baselines by a large margin on both success rate and mean distance to goal across diverse obstacle conditions.The comparisons assess the contributions of the dual-agent architecture and Humanoid-Object Decoupled Potential Fields over 20,000 trials in 20 scenes.Source: A. Simulation Experiments · p. 6, Table I · p. 6
  • On 50 unseen scenes, HOTICE averages 80.1% success with an approximately 70% floor and 0.27 m mean distance to goal.Success tends to fall and mean distance to goal rise as scenes become less similar to training scenes.Source: A. Simulation Experiments · p. 6, Figure 5 · p. 6
  • Performance on both metrics remains consistent across box, cylinder, and sphere objects.The shape-specific policies use the same 20 scenes as the main experiment.Source: A. Simulation Experiments · p. 6

B. Real-World Results

On a Unitree G1, HOTICE transported objects through side, overhead, and ground obstacles, with strong trial results and qualitative advantages over its ablations. The section also describes deployment setup and results across object shapes.

B. Real-World Results

  • Obstacle scenes are reconstructed with SLAM before transport, while inference uses robot localization, joint angles, and box pose estimated from AprilTags.The carried box occludes the G1’s single LiDAR sensor; the authors characterize this as a hardware limitation.Source: B. Real-World Results · pp. 6–7
  • HOTICE generalizes to cylinder and sphere objects with minimal shape-specific reward tuning.Table II reports three shape-specific policies distilled from teachers trained on the same 20 scenes.Source: Figure 6 · p. 7, Table II · p. 7
  • In real-world experiments across side, overhead, and ground obstacles, the proposed method succeeded in all six side-obstacle trials, with six trials conducted per obstacle type for each method.Source: B. Real-World Results · p. 7
  • Real-world rollouts show HOTICE performing effectively across obstacle types; the SGF-only policy collides more often, while the single-agent policy struggles with ground-obstacle hurdling.The single-agent policy also struggles to remain stable when stepping over the ground pipe.Source: Figure 7 · p. 7, B. Real-World Results · p. 7

VI. DISCUSSION AND LIMITATIONS

HOTICE’s limitations include unmodeled post-collision recovery, restricted object-manipulation strategies, and the exclusion of dynamic scenes. The authors also identify collision-tolerant navigation and active obstacle interaction as future directions.

VI. DISCUSSION AND LIMITATIONS

  • HOTICE does not model post-collision scenarios, and disturbances exceeding the robot’s recovery capacity cause it to fall.The method’s collision avoidance does not cover recovery from sufficiently large collision disturbances.Source: §VI · p. 8
  • The learned policy does not support more complex carrying strategies such as shoulder carrying or tucking an object against the hip.Source: §VI · p. 8
  • HOTICE does not address dynamic scenes with moving obstacles.The authors leave transportation in such settings for future work.Source: §VI · p. 8
  • The authors propose future work on acceptable collisions and actively moving obstacles, such as pushing small obstacles or doors.Source: §VI · p. 8

VII. CONCLUSION

HOTICE combines decoupled humanoid–object guidance fields, dual-agent whole-body control, and a generalist policy distilled from scene-specific teachers. Experiments in MuJoCo and on a Unitree G1 demonstrate reliable transport across clutter and object shapes, including previously unseen scenes and sim2real deployment.

VII. CONCLUSION

  • HOTICE is a whole-body humanoid learning framework for transporting objects through clutter that constrains the robot and payload from the ground, sides, and above.Source: §VII · p. 8
  • Decoupled guidance fields encode humanoid–obstacle and object–obstacle geometry, while dual-agent control and teacher distillation produce a generalist policy.Source: §VII · p. 8
  • MuJoCo and Unitree G1 experiments demonstrate reliable, collision-aware transport across diverse clutter and object shapes, including unseen scenes and sim2real transfer.Source: §VII · p. 8
Loading 2609.25363v1…