An arXiv paper compared the performance of a centralized commander, an auction-based consensus team, and independent LLM agents under five network conditions in a simplified perimeter-defense simulation. The consensus team had a 15% breach rate, the commander 32%, and independent agents 29%.
cuPIQP is a GPU batched differentiable quadratic programming (QP) solver based on a proximal interior-point method, with unified support for dense, sparse, and multistage problems. In benchmarks with up to 16,384 QPs per batch, it solves up to 18.8 times faster than Moreau and 4.26 times faster than eight-core PIQP; in an end-to-end learning task, training is accelerated by 4.6 times over QPAX and 7.8 times over Moreau.
AquaProbe proposes an underwater image adaptation method for frozen vision models: it measures how domain differences are distributed across encoder depths using paired clean images and physically rendered images, then allocates a fixed LoRA rank budget according to that distribution.
The authors propose ID-WM (Implicit Dynamics World Models), a hierarchical world model in which a high-level model samples feasible transitions on the state manifold and a low-level world model acts as an inverse dynamics solver to ground those transitions into precise actions. On three navigation and manipulation test suites, it improves over existing policies and diffusion planners by 38.5%, and by 135% over single-layer world model MPC in complex unseen long-horizon configurations.
RoboArchitect lets a coding model observe robot rollouts, diagnose failures, and iteratively revise the agent harness around a frozen policy. Without updating any model weights, it raises average success from 9.7% to 43.3% (4.5x) on unseen scenes and from 12.7% to 40.2% (3.2x) on unseen tasks.
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 study proposes Proof-Carrying Evolution (PCE), in which a language model provides both a controller and its certificate during search; only candidates that pass certificate verification can be returned, and the deployed artifact is the control law together with its certificate. The authors compare four language models with classical search and synthesis methods on speed-limited docking, maneuvering-target pursuit, and SO(3) attitude control. Certificates for linear problems are verified using exact rational arithmetic; on SO(3), floating-point verification is followed by exact re-verification of each reported champion. The best run of each method also receives a Lean 4 proof.
The authors propose Nexus-VLA, a hybrid policy learning framework that enables cross-embodiment knowledge transfer through end-to-end human-robot joint training, allowing human manipulation experience to directly contribute to robot policy learning. The framework first establishes hierarchical cross-embodiment semantic alignment from task intent to fine-grained interactions, then encodes multi-level shared knowledge into a shared manipulation latent representation used directly for continuous action generation. In RoboCasa-GR1, OOD settings, and real-robot experiments, Nexus-VLA achieved an 80.5% task success rate on RoboCasa-GR1.
A research team has released an open-source knee exoskeleton built from off-the-shelf components, whose biological torque controller is generated entirely by simulation, requiring no training data collected on the target device. The method uses a muscle-driven imitation policy to reproduce squatting motions from a public biomechanics dataset, then distills the knee joint torque into an estimator that relies only on onboard sensors.
The study proposes language-conditioned behavioral cloning for autonomous endoscope guidance, defining 6 clinically relevant guidance behaviors and collecting 235 subtask demonstrations. In preliminary phantom experiments, a single visuomotor policy achieved 100% success on focus, overview, and stop, 80% on search-tip and follow-tip, and 50% on horizon leveling. The authors note that the failure modes indicate more complex guidance tasks require temporal context.
The study extends two algorithms, CADENCE and its decentralized version DADENCE, to the dynamic version of POCGAGP (DPOCGAGP), adding two deployment-time perturbation settings, Agent Loss and World Change, for multi-agent systems to establish a fully covering communication network in unknown constrained spaces. On more than 4,500 benchmark test cases of varying scale and structure, the improved algorithms outperform another state-of-the-art benchmark algorithm. The work focuses on simulation benchmark evaluation and has not yet involved real-robot validation.
The study proposes an online method for learning closed-loop dynamics under a frozen locomotion policy, using a contaminated Gaussian model to distinguish predictable from unpredictable responses and combining a Gaussian process with random Fourier features to build an online learner that reduces bias from heterogeneous observations. Combined with model predictive control (MPC), the method predicts floating-base motion more accurately; in simulation and hardware experiments, end-effector tracking error was 0.01 m, and a finite-sample error bound for the online learning is provided.
The research team proposes Movens-Zero, a scalable understanding-generation-execution architecture that supports real-time closed-loop humanoid motion generation under text, speech, music, and video inputs. Its accompanying dataset, Movens-5M, is built from Internet videos through clip generation, quality filtering, semantic annotation, motion recovery, and physical refinement with humanoid retargeting. It contains 15 million human motion sequences and 5 million physically executable humanoid motion sequences, totaling over 10,000 hours.
Why it matters: This technical report presents a complete path for building a humanoid motion dataset from Internet videos and training a unified understanding-generation-execution model, offering a reference for data scale and architecture design.
The authors propose SimuGuard, an auditing plugin for LIBERO, RoboTwin, RoboCasa, and RoboDojo that identifies failures caused by the simulator rather than the policy. On RoboTwin, the authors trace these failures to the physics engine's collision detection: when geometry is 1 to 2 centimeters apart, the engine reports a penetration of about 2 centimeters, and the solver ejects the object at several meters per second because the distance algorithm discards triangles below a fixed area threshold.
Why it matters: The original traces failures in simulation benchmarks to a physics engine collision detection constant and provides a before-and-after success rate comparison, which can serve as a reference for teams doing simulation evaluation.
SketchMate has a vision-language model build a sketch model for insertion tasks, consisting of a few geometric primitives and two mating coordinate frames, which a planner and compliant controller then execute. In a strict kinematic test of 69 simulated insertion episodes, the method inserted 88% of parts at the assembled pose, while ReKep and MOKA using the same VLM achieved no successful insertions.
This paper proposes a flexible sensing manipulation framework that directly connects the structural dynamics of flexible links with sensing, kinematics, and control. The method first identifies the dominant vibration modes of each flexible link according to response convergence and excitation bandwidth criteria. Under the condition of full-rank modal observability, m retained modes require only m independent slope measurements and one reference IMU, and the continuous link deformation is reconstructed accordingly for flexible sensing forward kinematics and a PDE-based controller.
The arXiv paper "When the Past Shapes Action: Understanding Embodied Memory" defines embodied memory as the retention of decision-relevant information beyond the current observation, and organizes existing methods around six fundamental memory requirements and eight architectural categories.
ResMerge proposes a two-stage framework that learns mergeable residual policies atop a frozen pre-trained policy: residual reinforcement learning obtains task-specific corrections during adaptation without additional expert demonstrations, while parameter-space merging consolidates compatible residuals to reduce stored parameters.
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.
The authors propose the Direct Manifold Action Transformer (DMAT), which reframes bimanual action generation as directly modeling clean, observation-conditioned joint-space action chunks from stochastic interpolants, enabling single-step inference without iterative denoising or trajectory straightening.
RTPolicy is a multi-view visuomotor policy that uses view-aware feature fusion and visual compression to aggregate multi-camera information into a compact token representation, and uses a flow-matching action generator to reduce action-generation overhead. On the 50 tasks of the RoboTwin 2.0 single-task clean evaluation, it achieved an average success rate of 52.26%, higher than the strongest leaderboard baseline π0 at 46.4%.
The study applied five types of instructions—empty, neutral, meaningless, infeasible, and wrong-task—to six model checkpoints in the LIBERO-Goal simulation, holding the initial state fixed and varying only the instruction, to examine whether the VLA policy starts, which task it completes, and what the instruction determines.
BioAgent is an embodied agent for long-horizon laboratory protocols that acquires new skills by matching human-reviewed requirements with validated skills, selecting tool-based implementations for missing behaviors, and requesting human demonstrations as needed. On the second task, it reused 17 skills unchanged, modified 2 existing implementations, and added 5 new skill types; on two dual-arm laboratory tasks with 20 and 31 ordered skill invocations respectively, it achieved complete-task success rates of 66% and 58%, with each task run 50 times and autonomous recovery enabled.
The authors propose the MolmoSpaces-CommonSense benchmark, which builds a reasoning taxonomy grounded in developmental cognitive science and examines commonsense reasoning about occlusion, containment, support relations, semantically correct grasp points, spatial affordances, and hidden objects.
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 TaSP-KKT, which models the continuous-time limit of OmniRetarget's per-frame KKT conditions as an ODE on the robot configuration manifold and implements motion retargeting through forward discretization. The method requires only one weighted tangent-cone projection and a closed-form manifold update per step, avoiding repeated inner SQP iterations. In numerical experiments, retargeting quality is comparable to OmniRetarget but slightly lower, while running speed is about twice as fast.
The study proposes a joint-health-aware reinforcement learning framework that encodes actuator condition in a continuous joint-health vector and uses HealthNet to estimate each joint's health from recent observations and actions, allowing the policy to redistribute control effort across healthy and degraded joints.
HandMimicX proposes an object-centric framework for learning dexterous manipulation skills from egocentric human RGB-D videos. The method defines the reinforcement learning objective in terms of object motion, decoupling skills from specific hand trajectories, and imposes constraints on hand-object interaction states during grasping and manipulation to obtain more anatomically plausible hand configurations. Simulation and real-robot experiments show that the method learns efficiently from human videos and transfers skills to novel spatial configurations.
The authors propose HIPC, a real-time humanoid iterative planning and control framework that couples a generative model with motion tracking into a unified planning-and-control paradigm. HIPC uses goal-conditioned motion planning and trains with a canonical motion representation to avoid gimbal-lock singularities under extreme poses, while mitigating out-of-distribution shifts caused by real-robot state feedback through two strategies: deviation-decayed recovery and history re-anchoring. Experiments show that HIPC achieves the best goal-reaching accuracy among existing methods and exhibits robustness to physical disturbances and the ability to recover after falls.
The paper proposes the Hourglass architecture, which decomposes cross-embodiment motion planning into a body-agnostic world model and each robot's differentiable ego-dynamics template, coupled only through a camera ego-motion interface. On seven simulated robots across five motion types, this generalist policy achieves success rates 3.8–10.6 percentage points higher than imitation learning and diffusion policy baselines on the same world model, matches each robot's specialized policy without dedicated training, and comes within 1.4–2.8 percentage points of a parameter-conditioned MPPI teacher.
The authors propose LEXP, a framework for learning from physical experience through context-space adaptation that combines task-level planning with an experience module, reusing execution experience to generate contextual feedback while model weights and the action interface remain fixed. Across 12 real-world manipulation tasks, LEXP achieved 23/36 successes (63.9%), compared with 19/36 (52.8%) using memory alone and 2/36 (5.6%) for vanilla ChatGPT-6-Astra.
The research team proposes TRACE, an execution-time target-state risk gate that intervenes selectively while keeping the learned policy unchanged. They first used PPO to update only the planning head of a 0.6B VLA tracker; success and following rates improved across all three EVT-Bench settings, but the collision rate rose in distracted tracking scenarios.
TinyAstral proposes a two-stage method: first freeze an open-weight teacher model and optimize its harness through recursive self-improvement (RSI), then distill a 4B student model to replace the teacher within the same harness. On RoboTwin, harness evolution raised the teacher's success rate from 33.4% to 63.6%, and the final 4B student reached 62.8%; in measured deployment episodes, the student used about one-fifth of the teacher's accelerator time.
The authors propose the Capability-Conditioned Heterogeneous Graph Control Barrier Function (CC-HGCBF), which provides a unified safety representation for robots with differing kinematic and actuation capabilities in a shared second-order reference-point operational space. The method introduces the Radial Capability Field (RCF), which characterizes each robot's state-dependent finite-horizon reachable velocity set in a fixed dimension, enabling the shared graph barrier and decentralized policy to adapt to changes in state, kinematic constraints, and control limits without predefined robot identities.
The authors propose the Contact Imitation Interface (CIMI) framework, which uses contact geometry and fingertip RGB views to build a robot-free human demonstration interface and transfers skills to a robot hand through contact-alignment-first retargeting.
The authors propose a filter-agnostic pipeline that modulates the conservativeness of a safety filter at runtime using policy confidence: inside the filter's operational design domain (ODD), confidence is traded for performance; beyond the ODD, low confidence yields additional safety margin. Preliminary F1TENTH simulations show that the modulated filter combines the strengths of permissive and conservative filters.
The research team used a five-fingered robotic hand to grasp and insert RJ45 connectors, evaluating 281 imitation-learning policies spanning seven policy families in approximately 1,700 real-robot trials. Controlled ablations showed that observation and action representations had a greater impact than conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping, representation in the robot-base frame, and normalization within the point encoder each improved success by 70 to 100 percentage points. For the vision-language-action policy, switching to absolute action prediction improved insertion success by up to 70 percentage points.
Why it matters: Approximately 1,700 real-robot trials compared 281 imitation-learning policies, indicating that observation and action representations determine real-world success more than optimizer and sampling choices.
CaPtain proposes an open post-training pipeline that trains small open vision-language models (VLMs) into code-as-policy (CaP) robot agents, without relying on VLAs, frontier models, or privileged state at test time. The pipeline first performs supervised fine-tuning on synthetic trajectories generated by GPT-6 Astra, then applies multi-turn reinforcement learning, introducing a physics critic that reads simulator privileged state to assign credit to each request.
Desktop robots are becoming a new offshoot for embodied AI makers, focused on two functions: emotional companionship and productivity tools. The Letianpai desktop robot entered the market at a price of 1,799 yuan, the Lingtong Nian NIA-F01 offers companion interaction in a figurine form, and Elephant Robotics' myCobot series of small robotic arms requires secondary development. The article argues that consumer products still center on emotional companionship and have not yet truly delivered on boosting productivity, and that the future requires moving toward modularity and upgrading the ability to operate in the physical world.