Humanoid loco-manipulation requires whole-body interaction, which is challenging to collect at scale through robot demonstrations alone. HumanVerse provides a 500-hour dataset pairing egocentric video with synchronized body and hand motion to align viewpoints and refine motion references. This pipeline is used to validate methods by transferring supervision to robot tasks without target demonstrations. The work develops $\lambda_0$, a humanoid foundation model for whole-body loco-manipulation using robot-aligned human demonstrations. $\lambda_0$ shows significant improvement, achieving 90.0% success on SIMPLE tasks and 62.5% on real-world loco-manipulation.
Contact-UAN is a residual world model designed to correct actuator dynamics in humanoid control, addressing failures in transferring simulation-trained policies to hardware. It achieves this by preserving simulator physics while correcting actuator dynamics to match real-world rollouts. The method requires contact-consistent replay, using proprioception to initialize replays with consistent contact forces. Training with real-world walking data enables the UAN to generalize to tasks requiring accurate force and impulse regulation. This approach results in policies with improved torque efficiency, lower force-tracking error, and enhanced robustness.
A visuo-inertial framework is presented to estimate gait-phase and speed in wearable robots by fusing data from a chest-mounted camera and four lower-limb IMUs. This framework uses Anticipatory Terrain Perception (ATP) for terrain classification and Terrain-Aware Locomotion Mode Identification (TALMI) to determine locomotion context. These context estimates feed into a Gait Kinematic Encoder (GKE) to enable gait-phase and walking-speed estimation. The system achieves accurate gait-phase RMSEs and speed estimations with low error rates. Inertial models use sensor-fused Euler angles to minimize dependence on sensor mounting variations.
Robotic systems often face severe occlusion when trying to perceive a target, necessitating active perception to select viewpoints. Conventional methods use local next-best-view planning based only on immediate utility, which can be suboptimal for highly occluded targets. The Receding-Horizon Next-Best-View (RH-NBV) planner addresses this by evaluating sequences of future viewpoints. In simulations and physical experiments with occluded targets, RH-NBV achieved substantially higher target information recovery than single-step baselines. This improvement stems from multi-step reasoning, providing greater benefit for strongly occluded targets.
DexMimic is a generalist policy for dexterous hand-object tracking learned from human interaction datasets. It addresses the limitation of existing physics-based methods that train specialist policies for individual trajectories. DexMimic employs Value-Guided Reset Allocation to efficiently train the policy by adapting initialization to prioritize difficult phases and select optimal reset states. Experiments demonstrate improved tracking success, faster fine-tuning, and potential zero-shot generalization to new trajectories and objects. The method was validated on a real-world dexterous manipulation platform.
The research benchmarks vision-based 3D reconstruction techniques for perceiving the complex morphology of submerged sea ice in the Arctic. It evaluates Structure-From-Motion (SfM) and monocular depth estimation using data from the MOSAiC expedition and synthetic data from an Unreal Engine simulator. The method is quantitatively compared by registering the vision-based reconstructions with upward-looking multibeam sonar data. This work addresses the limitations of acoustic sensors by incorporating spectral and high-resolution textural information. The goal is to advance monitoring of under-ice habitats crucial for understanding ocean-ice-atmosphere heat exchange.
Deep space missions benefit from autonomous operation enabled by onboard learning algorithms. This work demonstrates training and inference of a thermal segmentation model directly on a 6U CubeSat's GPU in space. The model is then used for closed-loop attitude control by processing real-time thermal images to track the Moon. The research implemented ROS 2 Foxy and a secure file system to facilitate data transport and integrity. Over a year, twelve experiments advanced autonomous spacecraft systems under LEO constraints.
A collective learning framework is proposed to achieve unified robotic manipulation despite embodiment-dependent transition dynamics. The method separates shared task-space intent from embodiment-dependent transition dynamics using a two-stage approach. A predictive dynamics adapter is introduced to regularize the policy by forcing the latent state to predict future transitions, grounding the policy in actual robot behavior. Evaluations show the method prevents performance collapse on diverse simulated embodiments. The approach achieved over 93\% overall success in simulations, outperforming other baselines.