C's Robotics Paper Notes

ULTRA: Unified Multimodal Control for Autonomous Humanoid Whole-Body Loco-Manipulation arxiv.org/pdf/2603.03279 1. RL based retargetting 2. multi-modal command student distillation and finetuning so that it can switch between goal reaching vs reference tracking, mocap vs depth

HoMMI: Learning Whole-Body Mobile Manipulation from Human Demonstrations arxiv.org/pdf/2603.03243 for cross-embodiement transfer: 1. use look-at points in 3D space instead of direct head states 2. mask out arms

Watch Your Step: Learning Semantically-Guided Locomotion in Cluttered Environment arxiv.org/pdf/2603.02657 this works shows you can train a policy plus using a semantic map to avoid stepping on valuable things. this is why I believe locomotion should all be mapping based.

Training is multi stage. They first have a mimic policy and distill that to a base policy. Then they use AIP (AMP for interactions) to make the policy generalize instead of memorizing kinematic references. DF needs mocap, so they also distill this into vision policies.

arxiv.org/pdf/2602.21723 LESSMIMIC: Long-Horizon Humanoid Interaction with Unified Distance Field Representations use distance field (DF) as a representation for HOI. each link's traj can be defined by DF + grad of DF + vel_norm + vel_tangent. (tbc)

arxiv.org/pdf/2603.01126… Pro-HOI: Perceptive Root-guided Humanoid-Object Interaction trained with mimic + contact commands, but deployed with planner to replace the reference.

arxiv.org/pdf/2602.18164 The grandtour dataset, with tons of odometry methods benchmarked in tons of environments.

hero-humanoid.github.io/static/hero.pdf Learning Humanoid End-Effector Control for Open-Vocabulary Visual Loco-Manipulation For humanoid EE goal reaching, on top of WBC pose tracking, a neural model is trained to further correct small biases near the goal.

VIGOR: Visual Goal-In-Context Inference for Unified Humanoid Fall Safety arxiv.org/pdf/2602.16511 visual humanoid fall recovery 1) flat terrain human sparse reference 2) terrain-aware ref adjustment 3) keypoint tracking teacher 4) distilled to visual student without keypoint obs

arxiv.org/pdf/2602.15733… meshmimic: Geometry-Aware Humanoid Motion Learning through 3D Scene Reconstruction kinda like CRISP + OmniRetarget, but: 1) uses human edge and excludes depth edge in image to extract contacts, which is smart; 2) use polygonal primitives for clean scenes

Kids today are not citing wococo :(

arxiv.org/abs/2602.11758 Humanoid Agile Object Interaction Control via Dynamics-Aware World Model predict object states from proprioceptive obs, use it to transform object template pointcloud, and embed the transformed pcl as policy inputs.

@MagnusAnde16653 in the link?

apex-humanoid.github.io reward engineering + state machine engineering + nice progress-based rewards for reference-free humanoid agile climbing learning.

arxiv.org/pdf/2602.07845 Recurrent-Depth VLA Another way of VLM->diffusion or think-act model or whatever it is called; much faster and adaptive computation while reasoning in latent space.

arxiv.org/pdf/2602.07932 When predicting policy feasibility, can also use reconstruction header to detect OOD.

arxiv.org/pdf/2602.08370 humanoid badminton. motion tracking -> distillation into goal-conditioned policies -> AMP and physics-based finetuning.

HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control 1) models skateboard dynamics in simulation 2) phase-based design for multiple skills 3) human reference and kinematics-based guidance 4) sys-id sim2real No perception yet

arxiv.org/pdf/2602.01067 A Systematic Study of Data Modalities and Strategies for Co-training Large Behavior Models for Robot Manipulation Another paper from TRI on how to train your VLA