Paper Sweeps

Humanoid Papers from October 25 to October 31, 2026

16Locomotion11Motion Tracking12Loco-Manipulation8Whole-Body Control20Training Infrastructure1State Estimation / Perception

16 papers

Locomotion

Terrain, gait, robustness, and agile movement.

paper preview The effect of limb kinematics on the speed of a legged robot on granular media visual

01arXiv

The effect of limb kinematics on the speed of a legged robot on granular media

Tunes gait timing and leg motion to make a small legged robot move faster across sand.

Chen Li et al.

project clip

02arXiv

Differentiable Contact Dynamics for Legged Locomotion: Policy Learning and Contact-Implicit Trajectory Optimization

Lets quadrupeds learn gaits and trajectories faster with a differentiable contact model.

Gijeong Kim et al.

paper preview Continuous-Time Safe Planning and Control for Agile Quadrupeds visual

03arXiv

Continuous-Time Safe Planning and Control for Agile Quadrupeds

Plans collision-free Spot motion while respecting the robot's different forward and sideways speed limits.

Gage MacLin et al.

paper preview RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion visual

04arXiv

RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion

Lets a large-wheel biped climb steps without vision by reacting to wheel and ground contact.

Kyungmin Nam et al.

paper preview HC2Robot: Hyper-Connection Transformers with Predicted Foresight for Multi-Terrain Humanoid Parkour visual

05arXiv

HC2Robot: Hyper-Connection Transformers with Predicted Foresight for Multi-Terrain Humanoid Parkour

Predicts each foot's next landing from depth and body motion to improve humanoid parkour.

Wenhao Xu et al.

project clip

06arXiv

REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal

Switches a G1 between walking and crawling so it can pass through low spaces.

Shengzhang Xu et al.

paper preview A Bi-level Co-design Framework for Lunar Monoped with 2D Jump and Landing Control visual

07arXiv

A Bi-level Co-design Framework for Lunar Monoped with 2D Jump and Landing Control

Designs a lunar monoped's body and jump controller together so it can launch and land.

Sanjay Ramkumar et al.

paper preview A Trainable CPG Framework for Quadruped Locomotion visual

08arXiv

A Trainable CPG Framework for Quadruped Locomotion

Uses one compact controller to switch a quadruped among five walking gaits.

Xiyan Liu et al.

project clip

09arXiv

Remember Where You Walk: Learning Spatial Memory for End-to-End Legged Robot Exploration

Gives a Go2 spatial memory so it can explore without forgetting where it already walked.

Leon Magnus et al.

project visual A generalist policy for muscle-driven control across individual musculoskeletal models visual

10arXiv

A generalist policy for muscle-driven control across individual musculoskeletal models

Controls muscle-driven walking across many different human body models with one shared policy.

Bozhao Wang et al.

paper preview Learning to Exploit Passive Dynamics for Energy-Efficient Target Hopping of a Spring-Legged Quadcopter visual

11arXiv

Learning to Exploit Passive Dynamics for Energy-Efficient Target Hopping of a Spring-Legged Quadcopter

Coordinates rotors with a passive spring leg so a robot can hop toward targets using less power.

Ruigang Chen et al.

project clip

12arXiv

RAIBO-Wagon: Real-Time Navigation System for a Legged Robot Towing a Wagon

Lets a quadruped tow a wagon through narrow passages with wagon-aware walking and planning.

Insung Yang et al.

paper preview OmniQuad: a Wheeled-Legged Hybrid Robot with Omnidirectional Wheels visual

13arXiv

OmniQuad: a Wheeled-Legged Hybrid Robot with Omnidirectional Wheels

Combines mecanum-wheel feet with legs so a quadruped can drive in any direction and step over obstacles.

Francesco Iotti et al.

paper preview Compositional Diffusion Planning with Chunk Selection and Reachability Measure visual

14arXiv

Compositional Diffusion Planning with Chunk Selection and Reachability Measure

Joins diffusion-generated motion chunks only when the robot can safely reach the next chunk.

Chanyoung Park et al.

paper preview FORGE: Feasible Online Reduced-Order Reference Generation and Whole-Body Execution for Agile Humanoid Locomotion visual

15arXiv

FORGE: Feasible Online Reduced-Order Reference Generation and Whole-Body Execution for Agile Humanoid Locomotion

Generates feasible running and hopping references online, then tracks them on a G1.

Adrian B. Ghansah et al.

paper preview TACOS: Terrain-Aware COntact Supervision for Visual Footstep Planning visual

16arXiv

TACOS: Terrain-Aware COntact Supervision for Visual Footstep Planning

Learns from foot contact which terrain spots are unstable, then uses vision to choose safer footholds.

Jean-Michel Fortin et al.

11 papers

Motion Tracking

Reference motion, imitation, retargeting, and behavior priors.

project clip

17arXiv

ViT-H: Generative Video-To-Humanoid Control with Motion Tracking Feedback

Turns ordinary video into humanoid motion, using tracker feedback to keep the generated moves executable.

Zimo He et al.

paper preview Delay Randomization Widens the Delay Margin of Learned Humanoid Motion Trackers visual

18arXiv

Delay Randomization Widens the Delay Margin of Learned Humanoid Motion Trackers

Makes humanoid motion trackers tolerate command delays by training them with randomized lag.

Shaoqin Li et al.

paper preview HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model visual

19arXiv

HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model

Adapts a humanoid motion tracker to real robot dynamics using the robot's own movement experience.

Chao Chen et al.

paper preview MimicAll: Structured Embodiment Conditioning for Cross-Humanoid Whole-Body Motion Tracking visual

20arXiv

MimicAll: Structured Embodiment Conditioning for Cross-Humanoid Whole-Body Motion Tracking

Tracks whole-body motions across many humanoid shapes with one body-aware controller.

Jia Li et al.

project clip

21arXiv

Context-Aware Generation and Modulation of Expressive Motion Behavior using Multimodal Foundation Models

Uses the surrounding scene and human cues to choose and adjust expressive motions on a real robot.

Till Hielscher et al.

project clip

22Project

METATRON: A Promptable Whole-Body Execution Layer for Humanoids

Lets humanoids execute new timed body, speed, and style prompts with one frozen whole-body controller.

Chen Tessler et al.

project clip

23arXiv

GRAFT: Geometry-Aware Residual Adaptation for Foundation Trackers

Adds terrain-aware leg corrections to a frozen whole-body tracker so motions work on uneven ground.

Liang Li et al.

paper preview Mimetic Embodiment Transfer of Human-Humanoid Interaction from In-the-Wild Videos visual

24arXiv

Mimetic Embodiment Transfer of Human-Humanoid Interaction from In-the-Wild Videos

Turns videos of people interacting into coordinated humanoid motions such as hugging and handshakes.

Nghia Le et al.

paper preview Whole-Body Humanoid Control via Stochastic Motion Tokenization visual

25arXiv

Whole-Body Humanoid Control via Stochastic Motion Tokenization

Compresses humanoid motion into tokens so long generated sequences can be tracked by the robot.

Michael Drolet et al.

project clip

26Project

MorphMimic: Morphology-Conditioned World Model and Policy Sharing Across Humanoids

Shares one world model and motion tracker across several humanoid bodies and command styles.

Lizhi Yang et al.

paper preview SONIC-K1: Porting a Universal Humanoid Motion Tracker to a Small Humanoid on a Single GPU visual

27arXiv

SONIC-K1: Porting a Universal Humanoid Motion Tracker to a Small Humanoid on a Single GPU

Ports an open whole-body motion tracker to a small K1 humanoid using one GPU.

Kyrylo Kolesnichenko et al.

12 papers

Loco-Manipulation

Whole-body task contact: feet, torso, hands, and objects.

paper preview CoCarry: Contact-Support-Structured Retargeting for Dual-Humanoid Cooperative Carrying visual

28arXiv

CoCarry: Contact-Support-Structured Retargeting for Dual-Humanoid Cooperative Carrying

Coordinates two humanoids to carry one object while preserving their shared hand contacts and balance.

Ziyu Lin et al.

paper preview From Tabletop to Walking Bot: Deploying a VLA Model on a Legged Manipulator Robot visual

29arXiv

From Tabletop to Walking Bot: Deploying a VLA Model on a Legged Manipulator Robot

Gives a Go2 with an arm one control interface for walking, leaning, reaching, and following visual instructions.

Giovanni Minelli et al.

project clip

30arXiv

Learning Force-Reactive Humanoid Loco-Manipulation

Teaches a G1 to push and pull while walking by commanding hand forces together with whole-body motion.

Sihang Li et al.

project clip

31arXiv

Zero-HLM: Generalizable Humanoid Loco-Manipulation without Real Robot Data

Trains a G1 to walk to objects and manipulate them using simulation demonstrations instead of real robot data.

Dianyi Yang et al.

paper preview ThinkLocoManip: A Whole-Body Control Agentic Harness for Legged Loco-Manipulation visual

32arXiv

ThinkLocoManip: A Whole-Body Control Agentic Harness for Legged Loco-Manipulation

Gives an agent explicit body, driving, arm, and recovery tools for quadruped manipulation.

Nuo Chen et al.

project clip

33arXiv

HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience

Uses egocentric human experience to train a G1 to follow visual instructions with its whole body.

Yang Liu et al.

project clip

34arXiv

MOTIO: Agentic Skill Engineering for Reality-Grounded Humanoid Loco-Manipulation

Turns videos and scene assets into staged G1 skills, then fixes their contacts in physics.

SpatialtemporalAI

paper preview Networked Humanoid Intelligence: Communication, Coordination, Collaboration, and Social Interaction visual

35arXiv

Networked Humanoid Intelligence: Communication, Coordination, Collaboration, and Social Interaction

Coordinates networked humanoids for timed collaboration while each robot keeps its own motion safe.

Loc X. Nguyen et al.

project clip

36arXiv

HA-DiPO: Diffusion Policy Optimization for Heterogeneous Cooperative Transport

Coordinates quadrupeds with different bodies so they can carry one object together.

Xuebin Li et al.

paper preview Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization visual

37arXiv

Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization

Splits long walking-and-manipulation tasks into phases and trains one policy to complete each phase.

Simone Manni et al.

project clip

38arXiv

LLM-guided Planar Loco-Manipulation for Quadruped Robots

Uses a language planner to chain navigation and whole-body pushing skills on a Go2.

Georges Jetti et al.

paper preview ELEVATE: Agentic Controller Acquisition from Task Failure for Quadrupedal Manipulation visual

39arXiv

ELEVATE: Agentic Controller Acquisition from Task Failure for Quadrupedal Manipulation

Lets an agent train new elevation, reaching, and descent skills when a quadruped fails a task.

Shiqi Liu et al.

8 papers

Whole-Body Control

MPC, impedance, safety layers, and whole-body coordination.

paper preview HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training visual

40arXiv

HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training

Teaches a G1 humanoid to box by adding a compact planner above a frozen whole-body controller.

Xihan Ma et al.

paper preview Move, don't think: sample-efficient collision avoidance with lossy proximity sensors visual

41arXiv

Move, don't think: sample-efficient collision avoidance with lossy proximity sensors

Teaches an H1-2 humanoid to dodge nearby obstacles with compact proximity signals while staying upright.

Carson Kohlbrenner et al.

paper preview Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots visual

42arXiv

Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots

Shows when humanoid safety filters fail and tests safer ways to avoid collisions.

Tiffany Bao et al.

paper preview Active Reachability Learning for Humanoid Recoverability Monitoring visual

43arXiv

Active Reachability Learning for Humanoid Recoverability Monitoring

Predicts when a G1 can recover and interrupts unsafe walking before the robot falls.

Zeyuan Feng et al.

project clip

44Project

Ukemi-SafeFall: An Ukemi-inspired Injury-Aware Falling and Recovery

Teaches a humanoid to roll, brace, and stand back up after a fall while reducing hard impacts.

Tran Duc Anh et al.

project clip

45arXiv

PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control

Uses a diffusion policy's predicted future poses to steer a humanoid onboard without a separate tracker.

Lei Ye et al.

project clip

46arXiv

Collision-Aware Humanoid Whole-Body Control under Imperfect Tracking Targets

Corrects unsafe tracking targets around obstacles before sending them to a frozen Digit controller.

Mohitvishnu S. Gadde et al.

project clip

47arXiv

Humanoid Dodgeball: Learning Fast Dynamic Obstacle Avoidance Motions for Humanoids

Teaches humanoids to dodge thrown balls with fast directional whole-body movements.

Aoqian Zhang et al.

20 papers

Training Infrastructure

Benchmarks, datasets, rewards, simulators, and tooling.

project clip

48arXiv

RobustRLlib: A Unified Library and Benchmark for Robust Reinforcement Learning Algorithms

Provides tools for training robot policies that keep working when dynamics, sensors, actions, or timing change.

Xi Chen et al.

paper preview Regularised RL improves sim-to-real transfer visual

49arXiv

Regularised RL improves sim-to-real transfer

Trains ANYmal walking controllers that transfer to hardware and stay stable through pushes.

Danilo Vucetic et al.

project clip

50arXiv

PrefO2O: Zero-Shot Preference-Based Reinforcement Learning via Offline-to-Online Latent Preference Transfer

Transfers preferences between related tasks so robots can learn new behavior without collecting new labels.

Ruiqi Wang et al.

paper preview R^2E: A Role-driven Reward Evolutionary Framework for Automated Reward Function Design visual

51arXiv

R^2E: A Role-driven Reward Evolutionary Framework for Automated Reward Function Design

Uses specialized AI agents to design and improve reward functions for robot locomotion.

Shouhao Chang et al.

paper preview Conditional Flow Models for Offline Maximum Entropy Reinforcement Learning visual

52arXiv

Conditional Flow Models for Offline Maximum Entropy Reinforcement Learning

Learns rewards and dynamics from offline data to train simulated locomotion policies.

Hao Cheng et al.

paper preview Multi-Objective Reinforcement Learning Using Routed Ensembles and Trajectory Attribution visual

53arXiv

Multi-Objective Reinforcement Learning Using Routed Ensembles and Trajectory Attribution

Routes experience across several policies so robots can improve multiple goals without one goal taking over.

Zhenshuo Zhang et al.

paper preview Multi-Horizon Future Discovery for Physics-Informed Exploration visual

54arXiv

Multi-Horizon Future Discovery for Physics-Informed Exploration

Chooses exploration goals across several time scales while a physics model blocks impossible behavior.

Shuwei Wang et al.

paper preview Best Policy Tracking in Reinforcement Learning visual

55arXiv

Best Policy Tracking in Reinforcement Learning

Tracks which training checkpoint is truly best so walking experiments do not select a worse policy by accident.

Judith Echevarrieta et al.

paper preview Model Selection for Off-Policy Evaluation via Pairwise Bellman Validation visual

56arXiv

Model Selection for Off-Policy Evaluation via Pairwise Bellman Validation

Compares offline value models to choose better robot policies without running every policy.

Lars van der Laan et al.

paper preview Multi-Agent Reinforcement Learning of Co-Adaptive Hip Exoskeleton Control (MARL-Exo) for Metabolic Energy Reduction visual

57arXiv

Multi-Agent Reinforcement Learning of Co-Adaptive Hip Exoskeleton Control (MARL-Exo) for Metabolic Energy Reduction

Co-trains a human model and hip exoskeleton controller to reduce the effort needed for walking.

Yifei Yuan et al.

paper preview When Black-Box Robot Policies Fail: Online Sensor Diagnosis and Repair with a Transparent Surrogate visual

58arXiv

When Black-Box Robot Policies Fail: Online Sensor Diagnosis and Repair with a Transparent Surrogate

Detects which sensors have failed and repairs their inputs around a frozen quadruped walking policy.

Marc Ziegler et al.

paper preview Program-Level Adversarial Co-Adaptive Self-Play for Robotic Security visual

59arXiv

Program-Level Adversarial Co-Adaptive Self-Play for Robotic Security

Uses adversarial coding agents to harden G1 control software against attacks that disrupt walking.

Leo Benaharon et al.

paper preview Bridging Exploration and Loss of Plasticity in On-Policy Reinforcement Learning visual

60arXiv

Bridging Exploration and Loss of Plasticity in On-Policy Reinforcement Learning

Keeps PPO policies adaptable by preventing their exploration from collapsing during training.

Andrea Vasco Grieco et al.

paper preview Repulsive Population Policy Gradients for Behaviorally Diverse Reinforcement Learning visual

61arXiv

Repulsive Population Policy Gradients for Behaviorally Diverse Reinforcement Learning

Trains a population of policies to discover different useful gaits instead of all learning the same motion.

Saeed Hedayatian et al.

paper preview Freeze, Don't Clone: Fine-Tuning with Expressive Policies Without Retaining Offline Data visual

62arXiv

Freeze, Don't Clone: Fine-Tuning with Expressive Policies Without Retaining Offline Data

Fine-tunes expressive robot policies online without keeping their original offline training data.

Sami Nur Islam et al.

project clip

63arXiv

Training Controllable World Models via Action Influence Maximization

Collects informative robot actions so a controllable world model can learn from much less data.

Samuel M. Bateman et al.

paper preview MUSBRIC: LEGO-Compatible Pneumatic Muscle Bricks for Reconfigurable Musculoskeletal Robots visual

64arXiv

MUSBRIC: LEGO-Compatible Pneumatic Muscle Bricks for Reconfigurable Musculoskeletal Robots

Provides modular pneumatic muscles and a matching simulator for rapidly building walking robots.

Qiyao Wang et al.

paper preview Echo: Scaling Visual Reinforcement Learning with Off-policy Experience visual

65arXiv

Echo: Scaling Visual Reinforcement Learning with Off-policy Experience

Speeds up visual robot learning by letting a state-based policy gather many unrendered experiences.

Gagan Khandate

paper preview Is It the Curriculum or Just the Easy Start? Benchmarking Unsupervised Environment Design on Quadrupeds visual

66arXiv

Is It the Curriculum or Just the Easy Start? Benchmarking Unsupervised Environment Design on Quadrupeds

Separates real curriculum improvements from easy starts in quadruped terrain training.

Jakub Gilewicz et al.

paper preview Gym-Anything for Robotics: Scaling physical simulations with digital agents visual

67arXiv

Gym-Anything for Robotics: Scaling physical simulations with digital agents

Uses digital agents to build and verify robot simulation tasks, including G1 football training.

Pranjal Aggarwal et al.

1 paper

State Estimation / Perception

Terrain, state, slip, sensors, and local context.

project clip

68arXiv

SSIO: Self-Supervised Inertial Odometry

Learns body velocity from each robot's unlabeled IMU data to improve inertial odometry.

Yuheng Qiu et al.