paper preview
01arXiv
DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models
Noise-robust vision-based quadruped parkour using depth-denoising world models.
Paper Sweeps
Humanoid Papers from September 27 to October 3, 2026
32Locomotion13Motion Tracking22Loco-Manipulation2Whole-Body Control22Training Infrastructure
32 papers
Locomotion
Terrain, gait, robustness, and agile movement.
paper preview
02arXiv
Echo in the Steps: Learning Perceptive Humanoid Parkour with Gated Memory
Perceptive humanoid parkour with saliency-guided gated memory for foothold selection.
paper preview
03arXiv
MimicAgent: Quadruped Skills via Prompt-to-Trajectory Generation
Agentic framework for prompt-to-trajectory generation for quadruped RL training.
paper preview
04arXiv
TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion
Tactile learning framework using sole pressure sensing for safer humanoid foot-terrain interaction, phase-aware rewards, and sim-to-real alignment on Unitree G1.
paper preview
05arXiv
PASSAGE: Scaling Scene-Aligned Motion Learning for Perceptive Humanoid Traversal in Cluttered Environments
Perception-conditioned planner-tracker for humanoid traversal in cluttered scenes using human motion data and flow-matching.
paper preview
06arXiv
Generate, Track, Improve: Perceptive Multi-Skill Humanoid Locomotion with RL-Fine-Tuned Motion Generators
Two-layer perceptive humanoid locomotion: flow-matching motion generator from raw depth images plans whole-body trajectories; RL tracking policy follows them. RL fine-tunes generator on ter...
paper preview
07arXiv
Multifunctional Locomotion Control of Multi-Jointed BURs with Swimming and gait Capabilities
A multi-jointed underwater robot uses the same structure for swimming and legged locomotion.
paper preview
08arXiv
Passive-Dynamic-Walking-Inspired Dynamics Guidance for Energy-Efficient Humanoid Locomotion
Guide economical humanoid gait discovery with temporary passive-dynamic-walking-inspired conditions.
paper preview
09arXiv
Predictive Safety Curricula for Robust Legged Locomotion
Allocate training experience toward rare locomotion failures using predicted future safety cost.
paper preview
10arXiv
FutureRay: Control-Aligned Future Range for Agile Quadruped Navigation
Forecast clearance and encounter risk with footprint/reaction-braking constraints for quadruped avoidance.
paper preview
11arXiv
Humanoid Badminton: Learning Dynamic Racket Skills from Limited Human Motion Data
Hierarchical RL composes dynamic badminton skills from sparse imperfect human motion data.
paper preview
12arXiv
Model-Informed Safe Reinforcement Learning for Bipedal Locomotion via Step-to-Step Prediction
Combine analytical stepping safety certificates with learned bipedal locomotion.
paper preview
13arXiv
Multi-Terrain Mastery: A Comprehensive Controller for Bipedal Locomotion
One bipedal controller handles varied slopes, compliant ground and constrained stairs.
paper preview
14arXiv
Proprioceptive Force Estimation for Quadruped Locomotion and Human-Robot Interaction
One proprioceptive force estimate supports payload accommodation and leash-guided locomotion.
project clip
15arXiv
RECAST: Recasting Vision-Language Semantics into an Actionable Cost Map for Robot Navigation
Ground VLM judgments in a cost map that guides safe quadruped trajectory selection.
paper preview
16arXiv
Terrain-Aware Autonomous Planetary Exploration for Exteroceptive-Proprioceptive Mapping with Quadruped Scouts
Fuse terrain geometry and robot-terrain interaction to improve quadruped exploration and energetic planning.
paper preview
17arXiv
A Long-Legged, Direct-Drive Monopedal Robot Achieves Exceptional Jump Height
A lightweight direct-drive monoped jumps high without elastic energy storage.
project clip
18arXiv
Bundled Contact Gradients: Stabilizing Differentiable Simulation for Deployable Dynamic Tasks
Smooth stiff-contact gradients locally while retaining dynamics fidelity for deployable humanoid policies.
paper preview
19arXiv
Co-design of trajectory and morphology for a vertical jump-climbing robot
Co-design body geometry and motion for dynamic vertical bounding under adhesion limits.
paper preview
20arXiv
Quadruped Obstacle Avoidance and Footstep Planning with Distributed Low-cost Time-of-Flight Sensors
Distributed low-cost ToF sensing supplies near-field geometry for quadruped locomotion.
paper preview
21arXiv
Experience-Driven Continual Learning of Terrain Traversability for Quadruped Robots
Validation-gated continual learning maps pre-contact vision to uncertainty-aware, experience-derived traversability for a Unitree Go2.
paper preview
22arXiv
LocoWM: High-Precision Locomotion through World-Model-Guided Residual Adaptation
World-model-guided residual adaptation for high-precision legged locomotion with terrain leveling, acceleration compensation, and push recovery.
paper preview
23arXiv
DODGER: Safety-Guided Reinforcement Learning for Robot Navigation Among Dynamic Obstacles
Safety-guided RL for humanoid navigation among dynamic obstacles using CBFs without runtime filtering.
paper preview
24arXiv
NEXUS: Perceptive Whole-Body Control for Terrain-Adaptive Teleoperation
Perceptive whole-body controller for humanoid teleoperation across mismatched terrains using teacher-student learning on large paired motion corpus.
paper preview
25arXiv
Predictive Semantic Safety: From Visual Physical Reasoning to Safety-Critical Control
Predictive Semantic Safety framework using VLM for physical event prediction and conformal-calibrated safety filtering on quadruped.
paper preview
26arXiv
Active 3D weaves for load-bearing and damage-resilient locomotion
Shape-memory-alloy 3D woven walkers, crawlers, and four-legged robots combine limb-driven locomotion with extreme load bearing and post-compression recovery.
paper preview
27arXiv
ReCo: Response-Consistent Locomotion with Policy-Aware MPC for Legged Manipulation
Response shaping of RL locomotion policy plus identified closed-loop model enables policy-aware MPC for continuous world-frame EE tracking on Go2+X5.
project clip
28arXiv
Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics
VAPS uses viability predictors and a hierarchy of nominal, abort, and protective-fall policies to reduce hardware damage during humanoid flips on G1 and Oli.
paper preview
29arXiv
PROMO: Preference-conditioned Multi-Objective Reinforcement Learning for Quadrupedal Robots
PROMO conditions a single quadruped policy on runtime semantic preferences over tracking, stability, and efficiency while keeping locomotion priors fixed.
paper preview
30arXiv
Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch
EF-GAIfO estimates feasibility of state-only demonstrations from robot experience for imitation under embodiment mismatch, validated on simulation locomotion and real Unitree Go2 object-rea...
paper preview
31arXiv
Reactive Humanoid Multi-Contact Using Learned Stability Models
Reactive planner uses learned CoP region models to select hand bracing contacts that increase impulse resilience on the Alex humanoid, validated in simulation and hardware push recovery.
paper preview
32arXiv
The Effect of Gait Stability Based on Two Types of Impact Strategies for Two-Link Walking and Brachiating Robots
Numerical study of passive dynamic gait families in two-link walking and brachiating models under state-based and time-based impact switching.
13 papers
Motion Tracking
Reference motion, imitation, retargeting, and behavior priors.
paper preview
33arXiv
CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments
Shared G1 behavior latents transfer through a unified encoder to per-robot PPO humanoid trackers; supports motion, goals, rewards and flow-generated commands.
paper preview
34arXiv
Learning Expressive and Compositional Motion Representation via Spectral Skills
Spectral skills give hierarchical humanoid controllers a compact, compositional command interface.
paper preview
35arXiv
Track-and-Complete: Learning Humanoid Skills from a Single Failed Human Video
Learn useful humanoid skills from a failed attempt rather than requiring a successful demonstration.
project visual
36arXiv
GAE: General Action Expert for Real-Time Humanoid Teleoperation
Low-latency whole-body humanoid teleoperation across diverse human behaviors.
paper preview
37arXiv
HOI-Retarget: Contact-Centric Retargeting for Human-Object Interaction
Generate robot-compatible humanoid interaction references using contact-centric motion retargeting.
paper preview
38arXiv
SocialHumanoid: Towards Expressive Humanoid Behavior via One-Step Co-Speech Motion Generation
Generate speech-synchronized affective whole-body motion for continuous physical humanoid execution.
paper preview
39arXiv
Traceable Human-to-Humanoid Sign Language Benchmarking
Trace repairs and retargeting through a 20,648-sequence humanoid sign-language dataset and benchmark.
paper preview
40arXiv
ECHO-G: Embodied Co-speech Humanoid mOtion Generation
ECHO-G generates full-body co-speech motion for humanoids directly in robot space using joint audio-text conditioning.
paper preview
41arXiv
Dense Temporal Motion Retargeting for Legged Robots
Dense temporal motion retargeting (DTMR) jointly optimizes timing and control via sampling-based MPC for legged robots.
paper preview
42arXiv
TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion
TERRA reconstructs support terrain from scene-less motion, retargets it to a 354-muscle body, and trains one terrain-spanning locomotion policy.
paper preview
43arXiv
GestAdapt: Workspace-Conditioned Co-Speech Gesture Generation for Humanoid Robots
Speech-to-motion diffusion generates workspace-aware gestures in a shared representation, then retargets them under physical constraints to G1, GR3, and Reachy2 humanoids.
project visual
44arXiv
World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories
WMMs model dynamic 3D scenes as sparse SE(3) pose trajectories using flow-matching with per-token noise for any-to-any conditioning.
paper preview
45arXiv
Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of Generative Motion Models
Retargeting identifiability analysis and Source-Instance Fidelity diagnostic, with paired human-to-G1 and LAFAN1-to-six-humanoid studies beyond the abstract's animal-motion evaluation.
22 papers
Loco-Manipulation
Whole-body task contact: feet, torso, hands, and objects.
paper preview
46arXiv
Uni-VLaT: Whole-Body Tactile Adaptation of VLA Policies for Humanoid Loco-Manipulation
Whole-body tactile adaptation of VLA policies for humanoid loco-manipulation.
paper preview
47arXiv
DexWeave: Learning Dexterous Humanoid Loco-Manipulation from Human Demonstrations
Learning dexterous humanoid loco-manipulation from human demonstrations.
paper preview
48arXiv
Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation
PRISM expands a few human interaction videos into counterfactual training data for depth-conditioned humanoid pick-carry-drop policies.
project clip
49arXiv
EgoAlign: Bridging the Human-Humanoid Gap for Long-Range Loco-Manipulation
Construct robot-compatible action and state supervision from egocentric human demonstrations without physical-robot demonstrations.
paper preview
50arXiv
EgoHumanoid-V2: Human-to-Humanoid Transfer of Coordinated Whole-Body Skills for Loco-Manipulation
Transfer coordinated whole-body skills directly from egocentric human demonstrations to humanoids.
paper preview
51arXiv
EquivDP3: A SIM(3)-Invariant Point-Cloud Encoder for Data-Efficient Humanoid Loco-Manipulation
Geometric equivariance improves low-demonstration humanoid diffusion policies.
project clip
52arXiv
KPI: A Promptable Kernel for Physical Interaction on Humanoids
Add task-specific contact behavior between a trajectory source and an unmodified whole-body tracker.
project clip
53arXiv
OTRetarget: Joint Robot and Object Motion Retargeting via Optimal Transport
Surface-contact OT plus constrained IK jointly retarget G1 and object paths at native scale; 87% OMOMO contact Jaccard and 6/8 single-clip hardware trials.
paper preview
54arXiv
Humanoid Loco-Manipulation With Discrete VLA Model
Holo-M extends a language model vocabulary with discrete whole-body action tokens for humanoid loco-manipulation.
paper preview
55arXiv
WB-WAM: Heterogeneous Body-Hand Pre-training for Humanoid Loco-Manipulation
Joint video/action pre-training learns whole-body humanoid priors from heterogeneous partially annotated data.
project clip
56arXiv
HuGo: LLMs as Whole-Body Policy Code Designers for Humanoid Loco-Manipulation
Use an LLM to design closed-loop high-level humanoid policy code over a frozen whole-body controller.
paper preview
57arXiv
Praxis: Distilling Physical Interaction Priors from Egocentric Videos for Generalizable Whole-Body Manipulation
Combine one-shot human video priors with posture calibration and feedback for whole-body humanoid manipulation.
project clip
58arXiv
CEER2: Directional and Tunable End-Effector and Root Compliance for Humanoid Loco-Manipulation
Directional/tunable EE and root compliance via hierarchical RL for humanoid interaction.
paper preview
59arXiv
Locomotion-Grounded Humanoid Soccer: Task-Gated Reinforcement Learning of a Multi-Directional Kicking Library
A locomotion-first, task-gated Unitree G1 controller layers seven directional kicking skills onto an omnidirectional gait substrate.
project clip
60arXiv
Fiatlux: A Long-Horizon Benchmark for Humanoid Ladder Climbing and Light-Bulb Replacement
Open Isaac Lab benchmark coupling Unitree G1 ladder positioning and climbing with fragile-payload bulb replacement and disposal across one long-horizon episode.
paper preview
61arXiv
Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies
A broad benchmark study probes GPT-6 Astra's direct and controller-mediated robot actions, including dense humanoid locomotion and whole-body loco-manipulation.
paper preview
62arXiv
InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation
Test-time evolution of staged reward programs executed by a fixed object-aware FB behavioral foundation model on frozen humanoid controller.
project clip
63arXiv
HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution
18-task benchmark evaluating humanoid tool selection and execution across L0/L1/L2 levels and standard/decoy tool sets on Unitree G1.
project clip
64arXiv
MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending
MASkillBlender learns a shared decentralized high-level policy that blends fixed single-humanoid primitive skills for multi-humanoid coordination using only task-level rewards and local obs...
paper preview
65arXiv
Toward Humanoid Robots in Construction: A Teleoperation Feasibility Study
Teleoperation system for Unitree G1 using XR upper-body control and pedal locomotion for construction tasks.
paper preview
66arXiv
Towards a General Humanoid Loco-Manipulation Model via Egocentric Whole-Body Human Data Pretraining
λ0 policy pretrained on egocentric human data then adapted to Unitree G1 loco-manipulation via three-stage training.
paper preview
67arXiv
HumanoidTTT: Test-Time Capability Reuse for Efficient Humanoid Control
HumanoidTTT reuses validated full motions on Unitree G1 only from certified entry states and consolidates the store online.
2 papers
Whole-Body Control
MPC, impedance, safety layers, and whole-body coordination.
paper preview
68arXiv
AMBIT: Anticipatory Multimodal Body Recruitment for Bimanual Tracking on a Humanoid
Recruit pelvis and waist motion before underactuated humanoid arms saturate during bimanual tracking.
project clip
69arXiv
CompliantWBC: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets
Whole-body compliance on heavy humanoids with lower-body engagement under arbitrary-site forces.
22 papers
Training Infrastructure
Benchmarks, datasets, rewards, simulators, and tooling.
paper preview
70arXiv
HarnessPAI: An Evolving Harness for Physical AI
Model- and embodiment-agnostic harness for Physical AI including legged walking agent.
paper preview
71arXiv
ChronoSRL: Temporal Geometry for Self-Supervised Reinforcement Learning
Temporal-distance, arrival-hazard and occupancy critics accelerate self-supervised goal control; simulated Go2 gains, but PPO remains better and G1 walking fails.
paper preview
72arXiv
Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
Translate observation-only videos into inspectable temporal specifications for robot policy learning.
paper preview
73arXiv
CoHuB: A Simulation Benchmark for Multi-Humanoid Collaboration
Benchmark multi-humanoid collaboration under egocentric visual observations.
paper preview
74arXiv
GAUGE: Planner-Conditioned Active Calibration of Opaque Quadruped Velocity Interfaces
Calibrate an opaque quadruped velocity interface using planner-relevant uncertainty reduction.
paper preview
75arXiv
Graph-Based Simultaneous Path and Foothold Planning for Multi-Limbed Intra-Vehicular Robots in Space Stations
Jointly plan paths and discrete footholds for multi-limbed robots locomoting through space-station interiors.
paper preview
76arXiv
RoboCompiler: Graph-Native Compilation of Closed-Chain Robots for Consistent Modeling, Control, and Simulation
Compile consistent mechanical interfaces for closed-chain robot control and simulation.
paper preview
77arXiv
Beyond Policy Alignment: Closing the Planning-Learning Loop for Robot Control with Learned World Models
PL-MPC improves learned-world-model control at the critic, planner-terminal, and planner-to-policy interfaces.
paper preview
78arXiv
RoboAssist: Interactive Human-Humanoid Planning for Long-Horizon Surgical Assistance
Interactive humanoid surgical-assistance planning links evolving workflow evidence to executable loco-manipulation tasks and layered runtime safety.
paper preview
79arXiv
Game-Guided Skill Discovery through Self-Play for Playable Agent Control
Hierarchical self-play discovers a compact set of interpretable, human-playable motor skills, including locomotion primitives on a simulated Unitree G1.
paper preview
80arXiv
SynIL: Leveraging Synergy for Offline Imitation Learning from Imperfect Demonstration Datasets
SynIL turns local state-action synergy into dense surrogate rewards for offline imitation from mixed-quality locomotion demonstrations.
paper preview
81arXiv
Action Chunking Proximal Policy Optimization with Feedback Correction
Online PPO training method combines chunk-level action planning with per-step state feedback and is benchmarked on IsaacGym locomotion, including Humanoid.
paper preview
82arXiv
Affordance-Conditioned Decision Making: Bridging the Semantic-Spatial Gap in Zero-Shot Cross-Floor Vision-and-Language Navigation
PACE turns semantic staircase intent into executable affordance poses and closed-loop action sequences, materially improving cross-floor navigation and deploying on a Unitree Go2.
paper preview
83arXiv
Draft: A Parametric Tool for Robot Design Exploration
Data-grounded parametric compiler generates simulation-ready legged MJCF models and exposes morphology/control tradeoffs through common PPO locomotion curricula.
paper preview
84arXiv
MOCHA: Multi-Objective Co-Design using Hypernetwork Architectures
Multi-objective PPO hypernetworks jointly represent locomotion policies across Cheetah/Walker morphology and objective spaces, enabling Pareto and generalist design search from one trained...
paper preview
85arXiv
EIDA: Execution-Interface Dynamics Adaptation for Real-to-Sim-to-Real Robot Navigation
EIDA fits pose-increment and velocity-feedback models from target-platform data and uses them inside a lightweight simulator for navigation policy training.
paper preview
86arXiv
In CEM, a World Model Is Also a Proposal Mechanism
Crossed audit separating the scorer and proposal-generator roles of world models inside CEM planning on Walker and Cheetah tasks.
paper preview
87arXiv
ASENA: Self-evolving Agents for Embodied Navigation
ASENA couples coding agents to robot sensing, supervised execution and persistent workspace on Unitree G1; includes optional 4B monocular VLN policy ASENA-VLN.
paper preview
88arXiv
RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers
Benchmark evaluating coding agents on robotics development workflows including whole-body humanoid motion tracking on Unitree G1 using LAFAN1 clips in MuJoCo-Warp/MuJoCo-C.
paper preview
89arXiv
Diffusion Policy Improvement with Proposal-Conditioned Refinement Flows
PReFlow combines critic-based proposal selection from a behavior flow with a conditional refinement flow for offline RL policy extraction.
paper preview
90arXiv
FERPO: Forward Entropy-Regularized Policy Optimization
On-policy max-entropy RL that fits an entropy- and KL-regularized target via forward KL using SNIS without critic action derivatives.
paper preview
91arXiv
Semantic Map Sharing and Capability-Aware Coverage Planning for AI-Native 6G Robotic Coordination
Capability-constrained semantic coverage planning for aerial/quadruped teams, evaluated in Isaac Sim; terrain classes constrain both allocation and traversal paths.