project clip
01arXiv
Around the World: Unified Learned Locomotion on a 270 g Continuous-Rotation Quadruped
Lets a tiny quadruped walk upright or upside down and recover after landing.
Paper Sweeps
Humanoid Papers from October 4 to October 10, 2026
16Locomotion10Motion Tracking18Loco-Manipulation7Whole-Body Control23Training Infrastructure7State Estimation / Perception
16 papers
Locomotion
Terrain, gait, robustness, and agile movement.
project clip
02arXiv
Bidirectional Voronoi-biased Exploration Curriculum for Reinforcement Learning
Teaches an ANYmal to climb onto boxes without demonstrations by expanding practice from both the start and goal.
paper preview
03arXiv
Hierarchical Reinforcement Learning for Collision-Free Locomotion of an Underactuated Biped
Teaches an underactuated biped to steer around obstacles while keeping a stable walking rhythm.
paper preview
04arXiv
VAMPS: Visual and Motor Policies from Sampling-Based Planning
Turns planned examples into a reusable Go2 walking policy that transfers from simulation to hardware.
paper preview
05arXiv
ROOT: Discovering Rewards for User-Specified Embodied Behaviors
Searches for rewards that produce the gait a user asks for, then transfers the behavior to a Go2.
paper preview
06arXiv
Terrain-Dependent Intra-Cycle Leg Timing for Effective Locomotion on Granular Slopes
Changes leg timing within each stride so a hexapod can make better progress on loose slopes.
paper preview
07arXiv
PhoneBot: A Low-Cost Open Humanoid Robot Platform Reusing Smartphones
Builds a low-cost smartphone-powered humanoid that can walk and stand back up after falling.
project clip
08arXiv
Feeling Through the Load: Compliant Quadruped Locomotion under Payload Interactions
Lets a Go2 feel pushes through an unsecured payload and compliantly follow a person's guidance.
paper preview
09arXiv
Model-Based Geometry-Aware Generative Optimization for Constrained Locomotion Planning
Plans quadruped footholds and humanoid postures while respecting obstacles, balance, and body clearance.
paper preview
10arXiv
HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion
Learns compact, steerable humanoid walking styles from human motion and transfers them across several robot bodies.
paper preview
11arXiv
Energy-Efficient Gait Adaptation via Hierarchical Reinforcement Learning for Quadrupedal Locomotion Across Diverse Terrains
Automatically changes a quadruped's gait to save energy across different terrain and speeds.
paper preview
12arXiv
DAMP: Humanoid Locomotion via Denoised Belief Learning and Adversarial Motion Priors
Teaches a humanoid to walk over rough terrain using only onboard body sensing.
paper preview
13arXiv
Higher-Order Morphology Priors for Quadruped Reinforcement Learning Under Actuator Degradation
Helps a Go1 keep walking when several actuators weaken by reasoning about how its limbs connect.
paper preview
14arXiv
Teaching a Robot Dog New Tricks: Diverse Quadruped Skills via Combined Reinforcement and Imitation Learning with Adversarial Task Selection
Trains one Unitree B1 policy for 22 walking, posture, digging, and combined skills.
paper preview
15arXiv
Toward Lunar Legged Robots: Field Deployment Lessons at LUNA
Reports what lunar dust, lighting, terrain, and field operations demand from real legged robots.
paper preview
16arXiv
Walking on Roofs: Exploring the Potential of Walking Robots for Construction Work on Roofs
Tests a Go2 on sloped roof rigs to expose why it slips, loses balance, and fails to recover.
10 papers
Motion Tracking
Reference motion, imitation, retargeting, and behavior priors.
paper preview
17arXiv
KungfuAthleteBot: learning high-dynamic humanoid motion from video with unified robust recovery
Teaches a G1 humanoid athletic moves from video and lets the same controller recover after a fall.
paper preview
18arXiv
Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking
Helps humanoids follow whole-body motions while an active safety filter keeps them within safe limits.
paper preview
19arXiv
Continual Humanoid Motion Learning
Lets a G1 learn new motion skills one after another without forgetting earlier ones.
project clip
20arXiv
MotionSpaceFlow: Representation-Aware Flow Matching in Direct Motion Space
Generates human motions from text while allowing specific joints and moments to be constrained.
project clip
21arXiv
QF3: Fast Flow RL with Filtered Q-Gradients
Makes flow-based robot learning fast enough to train G1 walking and motion tracking from scratch.
paper preview
22arXiv
Beyond Retargeting: Low-Latency and Robust Humanoid Whole-Body Teleoperation with Learned Atomic Motion Primitives
Teleoperates a G1 with a library of full-body motion pieces that stays stable when human pose input is noisy or missing.
paper preview
23arXiv
BRACE: Adapting Whole-Body References for Force and Terrain Aware Humanoid Motion Tracking
Adapts humanoid motion references to uneven ground and hand forces so one G1 tracker can handle both.
paper preview
24arXiv
MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking
Finds where video-driven humanoid motions fail, then improves the training examples around those failures.
paper preview
25arXiv
Benchmarking Behavioral Steerability in Behavior Foundation Models
Tests whether humanoid motion models can obey combined style, constraint, and composition requests on a real G1.
paper preview
26arXiv
VioLA: Learning Generalist Humanoid Control Policies from Human Data
Learns reusable body and hand skills mostly from human data, then executes them through humanoid controllers.
18 papers
Loco-Manipulation
Whole-body task contact: feet, torso, hands, and objects.
paper preview
27arXiv
InterMimicGen: Scaling Humanoid Loco-Manipulation through Self-Evolving Motion Imitation
Builds a large library of successful humanoid object skills by repeatedly improving motions copied from people.
project clip
28arXiv
I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning
Gives one humanoid policy many object skills, including carrying, pushing, kicking, recovering, and chaining tasks.
project clip
29arXiv
Dataset-Free Compliant Humanoid Loco-Manipulation with Dynamic Online Posture
Lets a humanoid move and handle objects compliantly using only hand targets, without human motion recordings.
paper preview
30arXiv
CoDance: Learning Reactive and Compliant Human-Humanoid Interaction from Video
Teaches a humanoid to dance responsively with a person while keeping two-hand contact soft and stable.
paper preview
31arXiv
Exploiting Hierarchical Controller Structure in Contextual Parameter Learning for Humanoid Loco-Manipulation
Automatically tunes a humanoid's walking and whole-body controllers for pushing boxes of different weights.
paper preview
32arXiv
Transporting Unsecured Stacked Payloads with a Quadrupedal Robot via Multi-Objective Reinforcement Learning
Lets a Go2 carry unsecured stacked boxes while adjusting how carefully it balances the load.
paper preview
33arXiv
Humanoid Rickshaw Pulling: Whole-Body Locomotion under Coupled Wheeled Loads
Teaches a G1 humanoid to pull a loaded rickshaw weighing as much as 115 kilograms.
project clip
34arXiv
Safe Multi-Robot Collaborative Transport Using Density Functions
Coordinates two or three quadrupeds carrying one object safely through narrow passages.
project clip
35arXiv
Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating
Trains one humanoid policy to complete a long five-object rearrangement without stopping between stages.
paper preview
36arXiv
iGPC: Generative Motion Priors for Object-Aware Humanoid Interaction
Teaches a G1 to seek support, recover from pushes, and move objects using depth and LiDAR sensing.
paper preview
37arXiv
BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation
Provides 20 household tasks for comparing learned and agent-written humanoid manipulation policies on G1.
paper preview
38arXiv
Robust Nonprehensile Object Transport with Quadruped Robots
Carries loose objects on a quadruped's back while reducing sliding during turns and disturbances.
paper preview
39arXiv
Workhorse: Learning Robust Whole-Body Humanoid Loco-Manipulation from Human Data
Learns dynamic humanoid object skills from first-person human videos without robot-specific motion retargeting.
paper preview
40arXiv
HULK: Learning Whole-Body Forceful Loco-Manipulation for Humanoids
Teaches a G1 to walk while pushing, pulling, and carrying heavy loads with its wrists and torso.
paper preview
41arXiv
Co2Skill: Whole-Body Control via Skill Composition for Long-Horizon Human-Environment Interaction
Combines sitting, standing, climbing, stairs, navigation, and hand skills inside one humanoid controller.
paper preview
42arXiv
Being-M0.7: A Latent World-Action Model for Humanoid Robots
Turns more than 10,000 hours of human video and motion into executable G1 whole-body object skills.
paper preview
43arXiv
Humanoid World Action Model With Joint State--Action Generation
Predicts both the next humanoid action and the body state that action should produce.
paper preview
44Project
Reflex: Visually Grounded Reactive Humanoid Control
Lets a G1 see, step toward, catch, and retain boxes thrown by a person using onboard RGB-D.
7 papers
Whole-Body Control
MPC, impedance, safety layers, and whole-body coordination.
paper preview
45arXiv
Safe Streaming Flow Planning by Aligning Sampling Dynamics with Execution Dynamics
Plans safe movements continuously and filters each step before a legged robot executes it.
paper preview
46arXiv
Magnet-Aware Control of Legged Robots
Keeps a quadruped stable near strong magnets by modeling magnetic forces inside its controller.
paper preview
47arXiv
RACER: Residual-Adaptive Closed-Loop Estimation for Sampling-Based Planning in Wheeled-Quadruped Racing
Adapts a wheeled quadruped's racing controller online when real motion differs from simulation.
paper preview
48arXiv
LLA-MPPI: Rapidly Adaptive Whole-body Control of Legged Robots with GPU-Accelerated Parallel Simulations
Adapts whole-body quadruped control online to payloads, slippery ground, leg failures, and box pushing.
project visual
49arXiv
Precise SE(3) End-Effector Tracking in Whole-Body Humanoid Control
Tracks precise hand positions and orientations while a humanoid balances and walks.
paper preview
50arXiv
CSF: Contextual Safety Filtering for Motion Generators
Filters generated humanoid motions against the current scene before a G1 tries to execute them.
paper preview
51arXiv
FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems
Lets a G1 walk and avoid pushes with a learned safety filter that transfers to hardware.
23 papers
Training Infrastructure
Benchmarks, datasets, rewards, simulators, and tooling.
paper preview
52arXiv
Beyond Reward Hacking: Proxy Divergence Across Four Layers of a Staged Humanoid Learning Pipeline
Shows where humanoid training rewards and tests can accidentally reward the wrong behavior.
paper preview
53arXiv
Multi-Fidelity Policy Gradients Stabilize Data-Scarce Reinforcement Learning
Uses cheap practice runs to train H1 and ANYmal walking policies with less high-quality data.
paper preview
54arXiv
Co-design Gym: A Unified Benchmark for Embodiment-Policy Co-optimization
Benchmarks how robot body shape and control policies can be designed together across several legged tasks.
paper preview
55arXiv
Test-time Multi-agent Coordination by Decomposed Value Gradient Flow
Lets several robot controllers improve their coordination together at test time without retraining.
paper preview
56arXiv
A Bird's-Eye View of Iterative Reward Design
Compares ways to iteratively design robot rewards under the same training budget.
paper preview
57arXiv
P3: Persistent Particle Planning for Constrained Diffusion Control
Keeps useful candidate plans between control updates so a legged robot can replan faster under constraints.
paper preview
58arXiv
Reachability-Aware Diffusion Policy Optimization
Trains safer walking policies by predicting when a future action may first become dangerous.
paper preview
59arXiv
Direction-Conditioned Policies for Online Goal-Conditioned Reinforcement Learning
Guides joint-level robot control using the direction and distance to a goal.
paper preview
60arXiv
Preserving Unstable Modes Through Inverse Dynamics in JEPA World Models
Preserves unstable but controllable motions in learned world models for biped balance and planning.
paper preview
61arXiv
Revisiting Temporal Regularization for Smooth Control in Deep Reinforcement Learning
Produces smoother legged-robot actions without sacrificing as much task performance.
paper preview
62arXiv
The Robot Is Not Its Description: GaugeBench for Representation Robustness in Morphology-Aware Policies
Tests whether locomotion policies still work when the same robot body is described in a different coordinate convention.
paper preview
63arXiv
RFPO: Rectified Flow Policy Optimization for Embodied Control
Runs flow-based walking policies much faster on robot hardware while preserving performance.
paper preview
64arXiv
RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments
Benchmarks whether multimodal agents can directly operate robots across locomotion, balance, recovery, and manipulation tasks.
paper preview
65arXiv
ClimbLab: MATLAB Simulation Platform for Legged Climbing Robotics
Provides an open simulator for designing and testing legged climbing, gripping, footholds, and stability.
paper preview
66arXiv
LACE-CRAFT: Robot Co-Design with Actor Inheritance and Blackboard Collaboration
Co-designs robot bodies and locomotion policies by proposing and testing new body and reward variations.
paper preview
67arXiv
FlashNeRD: Performance-First Contact-Rich Neural Robot Dynamics
Speeds up learned robot dynamics for contact-heavy training and planning with quadrupeds and objects.
paper preview
68arXiv
LASER: Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL
Improves offline locomotion and manipulation policies while keeping them close to actions supported by the data.
paper preview
69arXiv
Q-Learning with Scalar Adjoint Matching
Fine-tunes flow-based robot policies efficiently across legged locomotion and bimanual manipulation.
paper preview
70arXiv
An Informational Curse of Horizon in Goal-Conditioned Policy Learning
Explains why distant goals can become hard for policies to distinguish and proposes a better training method.
paper preview
71arXiv
Learning Unknown Constraints without Unsafe Data via Optimality and Counterfactual Regularization
Learns hidden safety constraints from safe demonstrations without requiring examples of dangerous behavior.
paper preview
72arXiv
A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control
Focuses massive training runs near a quadruped's current skill level so one policy learns 13 terrain families.
paper preview
73arXiv
Informationally Decoupled Trajectory Design for Sim-to-Real System Identification
Designs complementary test motions that identify a Go2's physical parameters for better simulation transfer.
paper preview
74arXiv
Higher-Order Action Supervision Makes A Strong Policy Class
Improves low-data robot policies by supervising how their actions should change over time.
7 papers
State Estimation / Perception
Terrain, state, slip, sensors, and local context.
paper preview
75arXiv
VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning
Helps vision-based robot policies keep working when colors, textures, and camera views change.
paper preview
76arXiv
Self-Repairing Recurrent Ensembles for Real-Time Recovery from Distribution Shift
Lets walking policies repair themselves online after sensor readings shift or fail.
project clip
77arXiv
TACET: Context-Appropriate Acoustic-Social Navigation for Quadrupeds
Lets a Go2 choose socially appropriate paths while also controlling how loudly its feet hit the ground.
paper preview
78arXiv
What the Elevation Map Cannot See: Semantic-Aware Locomotion and Execution-Aware Navigation for Humanoid Robot
Helps a G1 notice floor hazards that height maps miss and correct navigation errors as it walks.
paper preview
79arXiv
CMP-IRRT*: A Perception-Assisted Height-Adaptive Planner for Quadruped Robots
Lets a Go2 take shorter routes by stepping over low obstacles and going around high ones using a camera-based height map.
paper preview
80arXiv
Taming an End-to-End Autonomous Driving Policy for Urban Navigation of Quadruped Robots
Adapts autonomous-driving ideas so a Go2 can plan sidewalk-aware urban routes toward a goal.
paper preview
81arXiv
TAPNAV: Humanoid Navigation through Tactile Active Perception
Lets a vision-denied G1 choose where to touch its surroundings and navigate despite uncertain position.