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.
Professor Lan Xuguang's team at Xi'an Jiaotong University has proposed AffordSim, which connects open-vocabulary 3D affordance prediction to simulation trajectory generation. Starting from a natural language task description, it locates an object's functional regions, then combines cuRobo motion planning to generate executable manipulation demonstrations.
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.
A team at the University of California San Diego proposed EgoThumb in an arXiv preprint, a wearable robotic thumb driven by a palm-side wide-angle camera and an imitation learning model that can autonomously grasp and stabilize objects, allowing users to complete tasks such as unscrewing bottle caps and stirring tea with one hand.
A paper proposes that how tactile data is represented may matter more than resolution. The authors pretrained a tactile encoder, LingBot-Touch, on various sensors ranging from 3x3 glove arrays to 240x240 optical fingertips, and tested it with 4 strategies on 26 real tasks.
Ken Goldberg's team at UC Berkeley and Spencer Huang, NVIDIA's head of robotics engineering, proposed an "agentic robotics" method that lets AI agents generate, test, and correct robot control programs in simulation without additional data collection.
Why it matters: The report describes a method for using multi-agent systems to generate, test, and correct robot programs in simulation, and presents real-world data from warehouse sorting.
VLMs can understand semantics: which object to grasp and where to put it. But how do we turn this knowledge into fluent and dexterous actions? On Tuesday at CoRL, Andy and Will will present Flow Reversal Steering: a method for passing VLM "prompts" to a diffusion VLA.
Come see the RLT presentation at CoRL on Tuesday! Charles will explain how to adapt VLA models with sample-efficient real-world reinforcement learning.
The paper analyzes the spectral mismatch between the high-rank transition dynamics of continuous environments and the low-rank bottleneck in forward-backward (FB) representations, and proposes temporal abstraction to mitigate this problem. Starting from the spectral properties of the transition operator, the authors show that temporal abstraction acts like a low-pass filter: it suppresses high-frequency components and reduces the effective rank of the successor representation, while preserving a formal bound on the value function error. Experiments show that this alignment is a key factor for stable FB learning at high discount factors. The work has been accepted to NeurIPS 2026.
The authors propose LeWAM, a bidirectional Transformer trained end-to-end on a decoder-free JEPA latent space that supports four modes: forward, backward, inverse dynamics, and policy prediction. Linear probe experiments show that LeWAM's latent space reads robot and object state better than the forward-only Le World Model, and that its degree of ignoring visual distractors is comparable to LeWM and clearly better than reconstruction-based WAM.
VOMMI is a portable demonstration collection and learning framework that connects portable RGB demonstrations to VLA post-training through offline trajectory reconstruction and online visual-motion conditioning. R2-VO refines trajectories using sparse geometric anchors, reducing the average absolute trajectory error of body and hand streams by 24.6%; a policy post-trained only on portable demonstrations has an 18.2% lower base velocity error than a policy trained on robot-collected demonstrations, with comparable end-effector translation accuracy.
The study constructs a geometric visual pretraining dataset containing only a plane, an object, and a cube hand, with trajectories generated automatically and scenes without textures or backgrounds, so that the policy is first exposed to the geometric relationship between the hand and the object. The authors evaluate ACT on three simulated robots with five manipulation tasks each and on three real-world robot tasks; in most combinations, fine-tuning from the geometric prior yields higher success rates in early training than training from scratch, and requires only a small number of task demonstrations. The results indicate that highly simplified geometric scenes can also provide an initialization that transfers across robots and to real-world tasks.
The authors propose LAG-Fusion, a latency-aware guidance fusion framework for asynchronous multimodal diffusion policy composition that allows modality-specific policies to run at their own inference rates and provide denoising guidance whenever output is available. To keep asynchronous composition consistent, the authors derive a reference-frame rebasing rule for diffusion variables under relative action representations, so that delayed guidance is aligned before fusion. In contact-rich manipulation, the authors compose a low-frequency vision policy with a high-frequency force policy; experiments with heterogeneous modality latency show better responsiveness and task performance than synchronous fusion and a purpose-designed force-aware baseline.
EvoPlan proposes a neuro-symbolic planning framework that uses symbolic validity and constraints learned from data as guarantees, and can run on locally deployed open-weight models.
NoContactNoWorries proposes a Transformer-based multimodal framework that fuses RGB-D vision with robot proprioception to infer binary contact states for hand-object interaction as a pseudo-tactile signal. The authors train a single contact prediction model on multiple objects and verify that this inferred signal can support downstream reinforcement learning agents in completing in-hand object reorientation, and that it generalizes to new objects. Both simulation and real-robot experiments validate the feasibility of inferring contact from vision and proprioception. The paper has been accepted to IROS 2026.
The authors propose DexTeleop-0, a bimanual dexterous teleoperation framework that uses fingertip tactile feedback to overlay robot-side contact corrections on human tracking commands, mitigating kinematic mapping errors and the inability of vision alone to regulate contact force. The method computes localized corrections in real time via contact-stiffness approximation and the kinematic Jacobian, and adds per-contact force tracking, soft regularization of predicted contact moments, and bounded hand-joint residuals.
The paper proposes that autoregressive policies can achieve real-time execution by adjusting the tokenization horizon and applying constrained decoding, thereby guaranteeing strict latency bounds and enabling multi-trajectory decoding. In simulated and real-world environments, the autoregressive policy significantly improves task completion speed over synchronous inference and consistently outperforms equivalent-level flow-matching policies. The authors note that autoregressive policies also have inherent advantages of faster convergence and better generalization in instruction following. The work was published at CoRL 2026.
MosaicIMU is a carrier-conditioned Mixture-of-Experts (MoE) pretraining and adaptation framework for generalizable neural inertial odometry. It composes carrier-specific expert features through a prototype-based router, decodes local velocity and uncertainty constraints, and integrates them with a history-aware EKF. For unseen domains, it freezes the pretrained base model and learns a lightweight expert residual branch; for edge deployment, it reuses the router to select online samples for incremental updates.
CLAW is a fully end-to-end self-supervised framework that jointly learns a world model and continuous latent action representations directly from videos without action annotations. It uses adversarial latent regularization to prevent the latent variables from encoding future observations, so that they capture action structure, and a diffusion model models environment dynamics as a distribution over future observations rather than a single deterministic rollout. The authors say this latent action world model supports imitation learning from observation and goal-directed planning, and outperforms existing methods in multi-task, multi-embodiment experiments.
The study proposes Learning-Augmented Robotic Automation, a hybrid system that integrates learned task controllers and a neural 3D safety monitor into conventional industrial workflows to automatically perform flexible cable insertion and soldering on a motor production line. Using less than 20 minutes of real-world data per task, the system ran continuously for 5 hours 10 minutes without physical fencing, producing 108 motors with a 99.4% pass rate on product-level quality inspection, a takt time close to that of human workers with smaller cycle-time variation, and solder-joint quality comparable to that of skilled operators.
Why it matters: The paper reports data from continuous operation of learning-augmented industrial automation on a real production line for 5 hours 10 minutes, with 108 motors and a 99.4% pass rate, serving as a practical reference for learned control entering manufacturing.
The authors propose the Visual Learned Distance (VLD) framework, which decouples perception learning from policy learning for image-goal navigation: first, a distance-to-goal predictor is trained self-supervised on about 3,000 hours of internet and simulated video; then, in simulation, a reinforcement learning policy is trained using privileged geometric distances perturbed by GeoNoise; at deployment, VLD predictions are used zero-shot.
Sketch-MoMa proposes a teleoperation scheme for mobile manipulators that uses user hand-drawn sketches as instructions, employing a vision-language model (VLM) to understand sketches overlaid on observation images and infer the drawn shapes and the robot's low-level tasks.
Pranjal Aggarwal and Sean Welleck propose in an arXiv paper to formulate embodied tasks as interaction with digital software: the software renders the physical environment as a screen image and provides controls for acting in it, and a computer-use agent (CUA) completes the task using screenshots and keyboard and mouse actions.
Augmentron is a geometry-aware data augmentation pipeline that generates reusable scene representations and renders them according to recorded camera poses, producing environment and workspace variations that are consistent across time and multiple views while preserving the robot, task objects, actions, and annotations.
Flex-RSI is a benchmark for evaluating agents' recursive self-improvement capability under a fixed budget, covering perception, reasoning, robot action, 3D reconstruction, visual creation, and infrastructure tasks. In each round, an executor improves the task solution, which is kept only if it scores higher on held-out validation data; an improver then revises the harness, including the executor and itself. Final solutions are scored on hidden test data.
Addressing the problem that PPO uses a fixed clipping interval for all samples and cannot distinguish extreme importance ratios from tolerable large changes, the authors propose Ordered Policy Optimization (OPO): within each minibatch, it sorts the importance ratios, selects the extreme ratios that have the greatest effect on the update, and assigns them adaptive per-sample clipping radii according to a preset tail profile. The objective remains first-order and requires only minibatch sorting and a modified policy loss.
The paper proposes agentic manipulation policies (AMP), in which a generalist AI model directly controls a robotic arm and adjusts its actions as the task progresses. The authors' reference implementation, AMP1, runs unchanged across single-arm, bimanual, and mobile manipulation.
Why it matters: The paper reports zero-shot success rates and real-world trial data for a generalist AI model directly controlling a robotic arm, which can be compared with specialist policies such as VLA and WAM.
The authors propose the Seeing What Matters benchmark to evaluate task-conditioned actionable spatial belief in simulated tabletop manipulation, covering 860 frozen evaluation cases across three scenarios—coffee workstation, drawer/cabinet, and microwave—which expand to more than 115,000 evaluation episodes.
SURE is a data-efficient vision-language-action (VLA) training framework that models spatial uncertainty with a lightweight transposed-convolutional probabilistic action decoder and uses Soft-DTW alignment supervision to accommodate temporal variation across demonstrations. On LIBERO, RoboTwin 2.0, and real manipulation tasks, SURE matches or exceeds baseline VLA manipulation success rates using only 10% of demonstration data or 20% of training steps.
ODEAction proposes a chunk-free action generation paradigm that replaces the action-chunk prediction common in VLA models with continuous action evolution. It characterizes the rate of change of the current action over evolution time using a transition field over action space; during execution, it takes the current action as the initial condition and recursively generates subsequent actions via ODE integration. The method is implemented as an action head and does not depend on a specific vision-language backbone architecture. Experiments on LIBERO, EBench, RoboTwin, and real robots show better task performance and trajectory quality than matched action-chunking baselines, with particularly notable gains on fine manipulation tasks.
The arXiv paper "Goodhart in the Planner" measured how sampling-based control with rewards learned from human preferences performs on the true objective: on HalfCheetah, Hopper, and Meta-World's drawer-open.
DKWBC proposes a dynamics-informed whole-body control framework that uses inverse-dynamics feedforward and joint PD feedback to supply tracking torques, while a residual policy learns the joint-torque corrections needed for balance. The authors state that this structure explicitly separates motion-related torque demands, simplifying the policy learning task, and that they validated in simulation and on real hardware that a single balancing policy can generalize across multiple motion datasets, maintaining tracking and balance under ground support, added payload, and physical interaction. The project homepage has been made public.
The research proposes having GPT-6 Astra read a robot's base policy weights and edit only 65 rows of token embeddings, raising the success rate on π0.5's 765 held-out paraphrased instructions from 86.3% to 92.0% (48 repairs, 4 regressions, paired p=1.31×10⁻¹⁰), all without gradient updates or new demonstrations.
This research proposes an inference-time obstacle-guidance framework for pretrained relative-action diffusion policies. It rolls out cumulative relative actions into a workspace trajectory, evaluates obstacle interactions along the trajectory with an SDF-based obstacle model, and during denoising propagates each obstacle response to all contributing action increments. In real-robot experiments across seven obstacle-constrained scenarios spanning four manipulation tasks, the method achieved 105/140 (75.0%) collision-free task successes, compared with 8/140 (5.7%) for absolute-action pointwise guidance. Ablations show that the temporally coupled form of relative actions yields larger gains under both obstacle models.
MDF++ is a visuo-tactile world action model that uses a diffusion backbone to learn latent representations of multisensory observations end-to-end, addressing the problem that existing world action models rely on frozen pretrained encoders whose representations are not well suited for control. The authors propose a multi-task learning strategy that leverages cross-modal correlations to learn task-relevant representations, avoiding the feature collapse caused by naive joint training. On contact-rich manipulation tasks including real-world bolt screwing and nut mating, it outperforms world action models built on frozen pretrained representations by up to 50%.
SUGARv2 proposes a unified framework for learning diverse visual humanoid loco-manipulation skills through interaction reference synthesis and a shared training pipeline. The method adapts reusable task templates to different spatial configurations to synthesize coordinated robot and object motion references at low cost, and uses reference-based imitation as a common learning objective across tasks, first training privileged teachers and then training visual policies through teacher distillation and reinforcement learning.
The study proposes a planner-verified counterfactual (CF) test: holding the grasp pose fixed, a solid obstacle is inserted into the approach corridor and its points are added to the model input, after which the planner re-derives the label, flipping it from feasible to infeasible. Across eight predictors spanning point clouds, voxels, depth, and object tokens, CF-pair AUROC falls from 0.920-0.996 in natural scenes to near chance (0.470-0.504), and this flip holds across four planners, two robotic arms, and a real-object obstacle.
The authors propose A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface, combining corrective language, spatial evidence, execution history, and action proposals into correction codes; during training, MetaBrain provides supervision and runtime guidance.
Why it matters: The framework uses code as a scaling interface, injects corrective language and execution history into the policy, and provides quantitative comparisons across benchmarks and real-robot tasks.
FastFunnels proposes a hierarchical multi-robot navigation method: a high-level reinforcement learning policy generates deformable elliptical envelopes in real time, while a low-level distributed model predictive control (DMPC) layer keeps robots within the envelopes and avoids collisions with each other.
The authors present RevoSim, a multimodal tactile simulator for the BrainCo Revo3 dexterous hand that jointly models 285 piezoresistive taxels and 5 visuotactile fingertips. The simulator can synchronously generate pressure, depth, RGB, and marker-displacement observations; on a single RTX 5090, its depth-conditioned marker-displacement configuration supports up to 4096 parallel environments. The authors trained in-hand manipulation and grasping policies entirely in simulation, laying a foundation for future sim-to-real transfer.