Apple's research team published the paper "Perceptive Humanoid Motion from Simulation" on arXiv, unveiling HALO, an autonomous control system for humanoid robots, and simultaneously open-sourcing the training and deployment code.
Gamma Intelligence, TeleAI and Unitree have jointly launched LocoManiLab, a unified evaluation framework for humanoid robot mobile manipulation. It covers five types of home spaces, 13 scenario instances and 234 tasks, with the Unitree G1 and Dex3 dexterous hands as the unified test platform.
ForPC is a force-aware whole-body control framework that regulates a humanoid robot's cart-pushing speed through two-stage reinforcement learning. The method first infers force latent variables and state latent variables from onboard observations without external force sensors, then uses a residual policy to coordinate arm, waist, and leg movements; in simulation, cart-pushing speed RMSE is reduced by up to 68.2% compared with a locomotion-only control baseline, and lateral error on multi-curve routes is reduced by up to 74.2%.
Apple's research team published the paper "Perceptive Humanoid Motion from Simulation" on arXiv, unveiling HALO, an autonomous humanoid robot control system, and validating it on the Unitree G1.
Why it matters: Apple's HALO splits task orchestration, motion generation, and tracking into three layers, and reports comparative results of 91% in unfamiliar environments and 70% on composite tasks, offering a look at its layered design and data synthesis pipeline.
Apple released the HALO humanoid whole-body motion control system, which procedurally generates 1,000 virtual residential scenes and synthesizes nearly 300 hours of paired data, enabling the Unitree G1 to master skills such as sitting on chairs, turning sideways to pass through a half-open door, crouching under a table, and pushing boxes.
Why it matters: Apple's HALO uses simulation-synthesized data to train humanoid whole-body skills, transfers them zero-shot to real scenes, and ablation experiments show the contribution magnitude of each module.
The authors propose soccerbot, a hierarchical reinforcement learning framework that distills a structured latent motion prior from large-scale human soccer motion capture data, constraining policy learning within a dynamically feasible motion manifold. A single skill-conditioned policy thereby learns moving, dribbling, trapping, and kicking in a unified way, while maintaining human-like motion and fluid transitions during skill switching. After robustness-oriented simulation design, the framework transfers zero-shot to the Unitree G1 humanoid robot, completing the full soccer skill set in real-world environments.
A research team at Seoul National University has proposed RISE (Responsibility-Induced Specialized Experts), a VLA framework that can collect demonstration data for locomotion and object manipulation separately and then learn jointly from them, enabling a robot to execute multi-stage tasks consecutively without full multi-stage demonstrations.
The research treats complete closed-loop robot policy implementations as reusable execution experience: a coding agent generates policy code from a few successful demonstrations and iteratively improves it using simulation feedback. Across four source tasks in RoboCasa, iterative optimization raised the success rate from 28.3% with initially generated code to 64.2% with validation-selected code, and the learning approach was validated on a physical Unitree G1.
ROOF proposes a reinforcement learning method with contact-supported geometric escape guidance that enables proprioception-only humanoid robots to first create clearance and then complete the rise under overhead obstacles. In simulation, the recovery success rate was 88.5% under plates that can only be lifted, compared with 18.9% without geometric guidance; it was 98.6% under free plates with randomized size, mass, and placement. After deployment on a Unitree G1, the policy successfully recovered in 95 of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements not seen in training.
The authors propose REA, a routed embodied agent runtime that lets a reasoning-focused System 2, an embodied router System 1, and reactive System 0 policies cooperate within a single runtime, with the core rule that a running policy never waits on the brain.
The authors propose ∇Mimic, which uses exact gradients from a differentiable simulator to train whole-body motion-tracking policies for humanoid robots, reaching a score of 0.82 in 20 minutes and 0.874 in 1.8 hours on a single GPU, whereas the world-model method takes 1.2 hours to reach 0.82.
Why it matters: The paper presents a complete scheme for training humanoid policies with exact gradients under differentiable simulation and open-sources the engine and code, making it easy to compare the efficiency of first-order and zeroth-order methods.
DiffWBC proposes replacing deterministic policies with a one-step diffusion model for general humanoid motion tracking, requiring only one forward inference per 50 Hz control step, achieving a local tracking error of 20.1 mm on the BONES benchmark, and deployed on a physical Unitree G1.
The study proposes REA (Routed Embodied Agent), a unified runtime that lets a reasoning-focused System 2, an embodied routing System 1, and reactive System 0 policies collaborate within the same framework, with the core rule that a running policy does not wait on the brain.
The study proposes a language-driven navigation framework that enables mobile robots to understand natural-language requests such as "go to the mailbox," identify the referenced object, estimate its position using RGB-D data, and generate a navigation goal that is then executed by the ROS 2 Nav2 navigation stack.
ActiveAlign proposes a framework that uses active perception to narrow the Real2Sim2Real gap, comprising two modules: ActiveAlign-Waving and ActiveAlign-Touching.
The authors propose HumiTransfer, a framework that converts in-the-wild human-human interaction videos into physically feasible human-humanoid interaction motions. The method first reconstructs 3D human motion with foot-ground contact, then transfers one participant to the target humanoid embodiment through interaction modeling, and uses post-retargeting optimization to ensure balance, contact, and collision avoidance.
REACH is an RL-based hierarchical framework that enables a Unitree G1 humanoid robot to traverse height-constrained environments through variable-height bipedal walking and crawling. Its policies regulate body height over complementary command ranges: 1.27–0.50m for bipedal walking and 0.50–0.20m for crawling, and a bidirectional transition policy connects crouched walking and crawling. The framework also combines a LiDAR planner that converts point clouds into collision-free three-dimensional paths and locomotion commands. Real-world experiments validated height and speed tracking, terrain traversal, mode switching, and confined-space traversal; the code and videos have been released.
GRAFT proposes a geometry-aware residual adaptation framework that, without changing the parameters of a frozen humanoid behavior foundation model (BFM) or its motion-reference interface, uses a lightweight adapter to apply terrain-conditioned corrections only to lower-body execution targets. The adapter is trained with task and foot-geometry rewards, is regularized so it produces no unnecessary corrections on flat ground, and is distilled from privileged terrain observations to onboard depth.
This study proposes using a stochastic motion tokenizer to generate concise discrete motion representations that can be fed directly into a whole-body tracking policy to execute dance, fighting, and walking motions on the Unitree G1. At matched capacity, the tokenizer reconstructs Unitree G1 motions better than FSQ and VQ-VAE. In real-robot demonstrations, each generated motion lasted 16.8 seconds, and a 35-second fighting-to-dance sequence was completed.
The study proposes extending PPO into phase-conditioned policy optimization, using a phase-conditioned critic, phase-gated rewards, and a multi-phase reset curriculum to train a single policy for long-horizon loco-manipulation tasks. The method adopts a decoupled hierarchical control architecture that separates low-level balancing dynamics from high-level phase execution, and validates multi-stage skills on real hardware on the Unitree G1 and a Boston Dynamics Spot equipped with a manipulator arm, while testing design choices through ablation experiments.
StepTrigger is a contact-state-triggered backdoor attack on VLM-powered legged robots. The trigger signal comes from the foot-ground contact pressure patterns produced when a Unitree Go1 quadruped walks across dense terrain, rather than from prompts or visible markers.
The study proposes DSMS, a contact-implicit, direct-simulation-based multiple shooting framework that converts kinematically feasible reference motions into dynamically feasible whole-body trajectories. It handles contact, friction, collision, and joint limits within a nonlinear program without presetting a contact schedule. Compared with existing retargeting methods, DSMS accelerates motion-imitation reinforcement learning training and yields policies with high success rates and low tracking error. The authors achieved zero-shot sim-to-real transfer on the Unitree G1, completing walking, jogging, contact-rich crawling, and a 180-degree jump-turn.
PrepTurn is a reinforcement learning framework for anticipatory maneuvering in continuous local navigation. It maps the current and next pose goals directly to joint actions, letting the upcoming goal influence how the robot approaches the current one. The framework uses a persistent-credit reward to encourage completing the entire goal sequence while in motion, and an automatic curriculum that gradually increases goal difficulty to stabilize turning behavior; simulation and onboard Unitree Go2 experiments show high sequential traversal completion rates and reduced stop-and-go motion.
Zero-HLM proposes a real-to-sim-to-real framework that turns smartphone captures and simulated source demonstrations into scalable training data for humanoid loco-manipulation. The study generated 4,800 physically validated demonstrations and achieved zero-shot transfer on three Unitree G1 tasks; GR00T N1.7 trained in simulation reached a mean full-task success rate of 66.7%, versus 46.7% with real-data training.
HuGe is a family of humanoid generalist vision-language-action (VLA) models pretrained on over 2,000 hours of egocentric human and robot data. It predicts latent motion tokens to drive a pretrained whole-body controller and outputs hand-control signals.
The study proposes HumanoidNDM, an online fine-tuning framework for humanoid motion tracking driven by a neural dynamics model trained on real robots, aimed at closing the dynamics gap between training and deployment. The method includes low-frequency reference perturbation data collection with an adaptive intensity mechanism, training a dynamics model based on proprioceptive history in Euclidean space combined with differentiable forward kinematics, and directly fine-tuning the policy through backpropagated model gradients.
The paper proposes an active reachability learning framework for learning the stochastic finite-horizon recoverability envelope of a given fallback policy, so that fallback can be triggered before an unrecoverable fall occurs. The method reframes data exploration as a reach-avoid problem, in which an uncertainty-guided explorer discovers unseen states while avoiding confidently unrecoverable regions, and combines temporal-difference and Monte Carlo objectives to improve in-distribution value prediction and uncertainty estimation. The authors evaluate on stochastic benchmarks and a 29-DoF Unitree G1 humanoid robot, reporting clear improvements in monitoring coverage and accuracy, and zero-shot transfer to real hardware.
Researchers developed a single control policy that lets Unitree's G1 complete ladder climbing, crawling, obstacle traversal, stepping stones, and recovery after falls by deciding the placement of nine different body parts. The robot can use its hands, elbows, knees, feet, and even pelvis for support rather than relying only on walking, achieving a 92.5% success rate on a simulated obstacle course, but how this flexibility performs in unpredictable real-world environments remains a bigger test.
ViT-H proposes a generation-plus-tracking video-to-humanoid control framework. It uses a diffusion model to generate humanoid reference motions directly from monocular video, skipping human motion reconstruction and retargeting, and uses motion-tracking results to post-train the generator with direct preference optimization. The authors also present an autoregressive variant that enables real-time humanoid video teleoperation with sub-400 ms latency. Experiments on the Unitree G1 show that ViT-H outperforms existing reconstruction-retargeting-tracking pipelines in motion quality and executability, and that performance improves with model size and with pre-training and post-training data scale.
The research team deployed a VLA model on a Unitree Go2 quadruped equipped with an AgileX Piper robotic arm, achieving closed-loop mobile manipulation, and reported choices such as model variant, action horizon, and demonstration conventions that made real-robot deployment feasible. For the pick-and-place task, expanding demonstration coverage reduced held-out end-effector error by 30%, and adding secondary skills such as pushing introduced new behavior modes without degrading baseline performance. Closed-loop real-robot experiments showed that dataset composition affects task success.
HuGe is a family of humanoid generalist vision-language-action (VLA) models pretrained on over 2,000 hours of egocentric human and robot data. It predicts latent motion tokens to drive a pretrained whole-body controller and outputs hand-control signals.
Researchers from Shanghai AI Lab, Zhejiang University, and Shanghai Jiao Tong University propose EgoExplorer, a wearable capture system that uses a head-mounted GoPro HERO13, 5 VIVE Ultimate Trackers, and MANUS Metagloves or Wuji gloves to record 100 hours of human activity data, aligning first-person video, full-body motion reconstruction, finger movements, and tactile observations.
The paper proposes the STAMP framework, which uses a "robot-to-human-to-robot" process to connect independently captured humanoid motions into continuous sequences. Its core is a learned alignment model: after motions are retargeted to a human skeleton, the model ranks candidate positions, headings, and transition durations for the subsequent motion, and a human motion prior then generates the transition under the selected alignment. The study evaluates the tracking success rate of the generated transitions on the Unitree G1 and ROBOTIS K1, and demonstrates long-horizon motion transitions in real-robot demos.
The authors propose a spatiotemporal retargeting method, STR, which uses a vision-language model to select relevant motion signals and body parts and identifies keyframes accordingly, allocating additional solver iterations to keyframes and higher interaction weights to the selected parts to reduce retargeting computational overhead.
The study trained 12 policies with the same recipe on a simulated Unitree G1, varying gravity, terrain geometry, and ground contact models (rigid or 3D resistive force theory, 3D RFT), and tested Earth-to-Moon fine-tuning and gravity randomization.
An arXiv paper presents LocoPilot, a single-operator bilateral teleoperation and data-collection framework deployed on a 12-DoF Unitree Go2 quadruped equipped with a 6-DoF AgileX Piper L manipulator, an Intel RealSense D435 wrist camera, and a D405 head camera.
EmbodyHOI runs a differentiable inverse kinematics (IK) solver for a specific humanoid embodiment inside the sampling loop of a frozen flow-matching hand-object interaction (HOI) trajectory generator, placing embodiment reachability constraints within the generation process rather than in post-hoc retargeting.
CoEvolve is a framework that represents team coordination as executable code over fixed robot motor skills. Through verifier-guided co-evolution, multimodal foundation models instantiate coordination strategies, physical verification evaluates each executed transition, and rejected behaviors trigger structured revisions starting from an accepted checkpoint.
A UC San Diego team completed two live pig gallbladder removal surgeries in an ordinary operating room using the teleoperated humanoid robot Surgie, with surgeons directing and controlling the procedures rather than the robot operating autonomously. Surgie is a modified commercial Unitree G1 that can quickly switch between standard surgical instruments and is smaller in size and footprint than conventional surgical robots such as Da Vinci. Participating surgeons reported intuitive control and reduced physical strain, but also noted overheating and the need for frequent repositioning.