RPL: Learning Robust Humanoid Perceptive Locomotion on Challenging Terrains

Humanoid perceptive locomotion has made significant progress and shows great promise, yet achieving robust multi-directional locomotion on complex terrains remains underexplored. To tackle this challenge, we propose RPL, a two-stage training framework that enables multi-directional locomotion on challenging terrains, and remains robust with payloads. RPL first trains terrain-specific expert policies with privileged height map observations to master decoupled locomotion and manipulation skills across different terrains, and then distills them into a transformer policy that leverages multiple depth cameras to cover a wide range of views. During distillation, we introduce two techniques to robustify multi-directional locomotion, depth feature scaling based on velocity commands and random side masking, which are critical for asymmetric depth observations and unseen widths of terrains. For scalable depth distillation, we develop an efficient multi-depth system that ray-casts against both dynamic robot meshes and static terrain meshes in massively parallel environments, achieving a 5-times speedup over the depth rendering pipelines in existing simulators while modeling realistic sensor latency, noise, and dropout. Extensive real-world experiments demonstrate robust multi-directional locomotion with payloads (2kg) across challenging terrains, including 20° slopes, staircases with different step lengths (22 cm, 25 cm, 30 cm), and 25 cm by 25 cm stepping stones separated by 60 cm gaps.

Flow Policy Gradients for Robot Control

Likelihood-based policy gradient methods are the dominant approach for training robot control policies from rewards. These methods rely on differentiable action likelihoods, which constrain policy outputs to simple distributions like Gaussians. In this work, we show how flow matching policy gradients -- a recent framework that bypasses likelihood computation -- can be made effective for training and fine-tuning more expressive policies in challenging robot control settings. We introduce an improved objective that enables success in legged locomotion, humanoid motion tracking, and manipulation tasks, as well as robust sim-to-real transfer on two humanoid robots. We then present ablations and analysis on training dynamics. Results show how policies can exploit the flow representation for exploration when training from scratch, as well as improved fine-tuning robustness over baselines.

VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation

A key barrier to the real-world deployment of humanoid robots is the lack of autonomous loco-manipulation skills. We introduce VIRAL, a visual sim-to-real framework that learns humanoid loco-manipulation entirely in simulation and deploys it zero-shot to real hardware. VIRAL follows a teacher-student design: a privileged RL teacher, operating on full state, learns long-horizon loco-manipulation using a delta action space and reference state initialization. A vision-based student policy is then distilled from the teacher via large-scale simulation with tiled rendering, trained with a mixture of online DAgger and behavior cloning. We find that compute scale is critical: scaling simulation to tens of GPUs (up to 64) makes both teacher and student training reliable, while low-compute regimes often fail. To bridge the sim-to-real gap, VIRAL combines large-scale visual domain randomization over lighting, materials, camera parameters, image quality, and sensor delays--with real-to-sim alignment of the dexterous hands and cameras. Deployed on a Unitree G1 humanoid, the resulting RGB-based policy performs continuous loco-manipulation for up to 54 cycles, generalizing to diverse spatial and appearance variations without any real-world fine-tuning, and approaching expert-level teleoperation performance. Extensive ablations dissect the key design choices required to make RGB-based humanoid loco-manipulation work in practice.

SPIDER: Scalable Physics-Informed Dexterous Retargeting

Learning dexterous and agile policy for humanoid and dexterous hand control requires large-scale demonstrations, but collecting robot-specific data is prohibitively expensive. In contrast, abundant human motion data is readily available from motion capture, videos, and virtual reality, which could help address the data scarcity problem. However, due to the embodiment gap and missing dynamic information like force and torque, these demonstrations cannot be directly executed on robots. To bridge this gap, we propose Scalable Physics-Informed DExterous Retargeting (SPIDER), a physics-based retargeting framework to transform and augment kinematic-only human demonstrations to dynamically feasible robot trajectories at scale. Our key insight is that human demonstrations should provide global task structure and objective, while large-scale physics-based sampling with curriculum-style virtual contact guidance should refine trajectories to ensure dynamical feasibility and correct contact sequences. SPIDER scales across diverse 9 humanoid/dexterous hand embodiments and 6 datasets, improving success rates by 18% compared to standard sampling, while being 10X faster than reinforcement learning (RL) baselines, and enabling the generation of a 2.4M frames dynamic-feasible robot dataset for policy learning. As a universal physics-based retargeting method, SPIDER can work with diverse quality data and generate diverse and high-quality data to enable efficient policy learning with methods like RL.

Adversarial Game-Theoretic Algorithm for Dexterous Grasp Synthesis

For many complex tasks, multi-finger robot hands are poised to revolutionize how we interact with the world, but reliably grasping objects remains a significant challenge. We focus on the problem of synthesizing grasps for multi-finger robot hands that, given a target object's geometry and pose, computes a hand configuration. Existing approaches often struggle to produce reliable grasps that sufficiently constrain object motion, leading to instability under disturbances and failed grasps. A key reason is that during grasp generation, they typically focus on resisting a single wrench, while ignoring the object's potential for adversarial movements, such as escaping. We propose a new grasp-synthesis approach that explicitly captures and leverages the adversarial object motion in grasp generation by formulating the problem as a two-player game. One player controls the robot to generate feasible grasp configurations, while the other adversarially controls the object to seek motions that attempt to escape from the grasp. Simulation experiments on various robot platforms and target objects show that our approach achieves a success rate of 75.78%, up to 19.61% higher than the state-of-the-art baseline. The two-player game mechanism improves the grasping success rate by 27.40% over the method without the game formulation. Our approach requires only 0.28-1.04 seconds on average to generate a grasp configuration, depending on the robot platform, making it suitable for real-world deployment. In real-world experiments, our approach achieves an average success rate of 85.0% on ShadowHand and 87.5% on LeapHand, which confirms its feasibility and effectiveness in real robot setups.

BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning

Building Behavioral Foundation Models (BFMs) for humanoid robots has the potential to unify diverse control tasks under a single, promptable generalist policy. However, existing approaches are either exclusively deployed on simulated humanoid characters, or specialized to specific tasks such as tracking. We propose BFM-Zero, a framework that learns an effective shared latent representation that embeds motions, goals, and rewards into a common space, enabling a single policy to be prompted for multiple downstream tasks without retraining. This well-structured latent space in BFM-Zero enables versatile and robust whole-body skills on a Unitree G1 humanoid in the real world, via diverse inference methods, including zero-shot motion tracking, goal reaching, and reward optimization, and few-shot optimization-based adaptation. Unlike prior on-policy reinforcement learning (RL) frameworks, BFM-Zero builds upon recent advancements in unsupervised RL and Forward-Backward (FB) models, which offer an objective-centric, explainable, and smooth latent representation of whole-body motions. We further extend BFM-Zero with critical reward shaping, domain randomization, and history-dependent asymmetric learning to bridge the sim-to-real gap. Those key design choices are quantitatively ablated in simulation. A first-of-its-kind model, BFM-Zero establishes a step toward scalable, promptable behavioral foundation models for whole-body humanoid control.

ResMimic: From General Motion Tracking to Humanoid Whole-body Loco-Manipulation via Residual Learning

Humanoid whole-body loco-manipulation promises transformative capabilities for daily service and warehouse tasks. While recent advances in general motion tracking (GMT) have enabled humanoids to reproduce diverse human motions, these policies lack the precision and object awareness required for loco-manipulation. To this end, we introduce ResMimic, a two-stage residual learning framework for precise and expressive humanoid control from human motion data. First, a GMT policy, trained on large-scale human-only motion, serves as a task-agnostic base for generating human-like whole-body movements. An efficient but precise residual policy is then learned to refine the GMT outputs to improve locomotion and incorporate object interaction. To further facilitate efficient training, we design (i) a point-cloud-based object tracking reward for smoother optimization, (ii) a contact reward that encourages accurate humanoid body-object interactions, and (iii) a curriculum-based virtual object controller to stabilize early training. We evaluate ResMimic in both simulation and on a real Unitree G1 humanoid. Results show substantial gains in task success, training efficiency, and robustness over strong baselines. Videos are available at https://resmimic.github.io/ .

OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction

A dominant paradigm for teaching humanoid robots complex skills is to retarget human motions as kinematic references to train reinforcement learning (RL) policies. However, existing retargeting pipelines often struggle with the significant embodiment gap between humans and robots, producing physically implausible artifacts like foot-skating and penetration. More importantly, common retargeting methods neglect the rich human-object and human-environment interactions essential for expressive locomotion and loco-manipulation. To address this, we introduce OmniRetarget, an interaction-preserving data generation engine based on an interaction mesh that explicitly models and preserves the crucial spatial and contact relationships between an agent, the terrain, and manipulated objects. By minimizing the Laplacian deformation between the human and robot meshes while enforcing kinematic constraints, OmniRetarget generates kinematically feasible trajectories. Moreover, preserving task-relevant interactions enables efficient data augmentation, from a single demonstration to different robot embodiments, terrains, and object configurations. We comprehensively evaluate OmniRetarget by retargeting motions from OMOMO, LAFAN1, and our in-house MoCap datasets, generating over 8-hour trajectories that achieve better kinematic constraint satisfaction and contact preservation than widely used baselines. Such high-quality data enables proprioceptive RL policies to successfully execute long-horizon (up to 30 seconds) parkour and loco-manipulation skills on a Unitree G1 humanoid, trained with only 5 reward terms and simple domain randomization shared by all tasks, without any learning curriculum.

LeVERB: Humanoid Whole-Body Control with Latent Vision-Language Instruction

Vision-language-action (VLA) models have demonstrated strong semantic understanding and zero-shot generalization, yet most existing systems assume an accurate low-level controller with hand-crafted action "vocabulary" such as end-effector pose or root velocity. This assumption confines prior work to quasi-static tasks and precludes the agile, whole-body behaviors required by humanoid whole-body control (WBC) tasks. To capture this gap in the literature, we start by introducing the first sim-to-real-ready, vision-language, closed-loop benchmark for humanoid WBC, comprising over 150 tasks from 10 categories. We then propose LeVERB: Latent Vision-Language-Encoded Robot Behavior, a hierarchical latent instruction-following framework for humanoid vision-language WBC, the first of its kind. At the top level, a vision-language policy learns a latent action vocabulary from synthetically rendered kinematic demonstrations; at the low level, a reinforcement-learned WBC policy consumes these latent verbs to generate dynamics-level commands. In our benchmark, LeVERB can zero-shot attain a 80% success rate on simple visual navigation tasks, and 58.5% success rate overall, outperforming naive hierarchical whole-body VLA implementation by 7.8 times.

Learning Gentle Humanoid Locomotion and End-Effector Stabilization Control

Can your humanoid walk up and hand you a full cup of beer, without spilling a drop? While humanoids are increasingly featured in flashy demos like dancing, delivering packages, traversing rough terrain, fine-grained control during locomotion remains a significant challenge. In particular, stabilizing a filled end-effector (EE) while walking is far from solved, due to a fundamental mismatch in task dynamics: locomotion demands slow-timescale, robust control, whereas EE stabilization requires rapid, high-precision corrections. To address this, we propose SoFTA, a Slow-Fast TwoAgent framework that decouples upper-body and lower-body control into separate agents operating at different frequencies and with distinct rewards. This temporal and objective separation mitigates policy interference and enables coordinated whole-body behavior. SoFTA executes upper-body actions at 100 Hz for precise EE control and lower-body actions at 50 Hz for robust gait. It reduces EE acceleration by 2-5x relative to baselines and performs much closer to human-level stability, enabling delicate tasks such as carrying nearly full cups, capturing steady video during locomotion, and disturbance rejection with EE stability.