The paper proposes the MAP (Masked Action Space) interface, which unifies spatiotemporal references such as keypoints, flows, orientations, and key poses as partial constraints in the same temporal Cartesian action space, and models controllable manipulation as an action inpainting problem.
The study proposes GR-VIP, a goal-conditioned visuomotor policy that uses a visual encoder pretrained for relative camera pose estimation, conditioning action prediction on dense relational features before pose regression to preserve fine-grained cross-view information for learning corrective actions from demonstrations. In real-world docking tasks, GR-VIP achieves high docking success rates and millimeter-level terminal accuracy in both the training environment and unseen environments, with transfer requiring no additional training. Supplementary experiments show the policy can align with non-docking targets and complete docking under the low-light conditions tested.
An arXiv paper proposes the Hardware-as-Regularization method, which uses a low-cost passive joint module (an adjustable mechanical clutch plus a 3D-printed spiral spring) to pull each joint toward a user-defined neutral posture, so that the passive restoring force acts as a physical prior for joint-space L2 regularization.
MambaAct proposes replacing the Transformer-based action expert in π0.5 with a compact Mamba action expert while retaining its pretrained vision-language prefix encoder.
Ego2Contact learns visual-tactile priors from human interactions and transfers them to robot hands. Its predictor fuses temporal scene context with hand-local visual features to infer bimanual tactile fields from RGB video, and anatomical region correspondences align varying robot tactile layouts with the predictor's canonical grid.
SenseSwarm runs a lightweight 709K-parameter map predictor on the swarm's shared occupancy map and folds the predicted structure back into a stigmergy-based guidance field, fine-tuning it on the swarm's own multi-front observations; a single-robot predictor fails on such views.
A CoRL 2026 paper proposes SpatialBEV, which queries bird's-eye-view (BEV) features around the predicted state at each step and conditions the next prediction on current geometric and semantic cues, in order to capture the complex interactions between terrain and vehicle in off-road scenes.
An arXiv paper proposes AnchorQ, which uses flow reversal to map reference policy behaviors into state-conditioned latent anchors, samples particles near the anchors, and applies value-guided sampling to achieve coarse-to-fine policy improvement. The analysis shows that anchors reduce identification complexity but introduce support bias; particle refinement trades this bias against action-class complexity, and critic accuracy determines the reliability of KL-regularized value-guided selection.
DefFlow is a goal-conditioned manipulation framework for deformable objects. By jointly predicting the goal-conditioned 3D point flows of the robot and the deformable object from demonstrations, it captures manipulation decisions such as grasping, releasing, and re-grasping, as well as the continuous fine motions required to reach the goal. Across rope, granular media, and cloth settings with diverse initial and target configurations, DefFlow clearly outperforms goal-conditioned action prediction and robot-flow-only baselines on success rate and goal-distance metrics, and it was deployed on a real robot using real point cloud observations and unseen target configurations.
An arXiv paper proposes using Gauss's principle of least constraint to formulate simultaneous force and motion control of a robot manipulator as a quadratic programming (QP) problem in generalized coordinates (joint space), so that inequality constraints such as joint limits, collision avoidance, and singularity avoidance can be incorporated directly into the solution. Simulations were run in MuJoCo and compared with an operational-space (task-space) method; the results show that the method can simultaneously maintain the desired contact force, reduce trajectory tracking error, and satisfy the inequality constraints. The paper is 8 pages with 5 figures and has been submitted to IEEE ICRA 2027.
PASS-VLN is a passability-aware benchmark and model for local-command vision-and-language navigation (VLN). It holds the scene, instruction, start pose, and reference route fixed while varying only the robot body width, requiring the robot to reach the target when the passage is feasible and to stop early when it is blocked.
This work proposes a visuomotor imitation learning pipeline for pedicel-targeted strawberry harvesting that unifies perception, motion planning, and contact manipulation: data is collected from human teleoperation demonstrations, an Action Chunking with Transformers (ACT) policy with a ResNet-50 visual backbone is trained, and the policy is deployed autonomously on a Jetson AGX Orin edge platform at 15 Hz.
We propose Shape-Aware Neural Time Fields, a physics-informed neural motion planning framework that learns a reusable travel-time (cost-to-go) representation for rigid-body motion, explicitly distinguishing the different geometric constraints on translation and rotation, and pairs it with a sampling-based planner.
The arXiv paper "Spatial Memory by Agents, for Agents" investigates letting coding agents automatically design robot spatial memory: under a fixed development budget, a Designer Agent writes and iteratively revises memory construction and retrieval code, with feedback from a fixed Query Agent; the resulting implementations are frozen and then used to build memory from held-out observation histories for evaluation.
A simulation study compared the speedup effects of two pipelines for multi-robot drawing: single-token (component-level) allocation and shared-token (decomposed) allocation, covering 385 drawings and team sizes of 1 to 8 robots. The results show diminishing speedup gains under single-token allocation; the shared-token pipeline achieved higher mean speedups and lower idle-time shares at larger team sizes, but it did not shorten completion time for every drawing.
FCTS is a failure-conditioned terrain curriculum framework that uses localized traversal failures to automatically generate the terrain a policy should train on next. Each failure's location, termination mode, robot state, and surrounding terrain geometry jointly condition the generation process, and Monte Carlo Tree Search explores alternative curriculum trajectories under a fixed training budget.
The paper proposes Embodiment-Classifier Guidance, which trains a lightweight embodiment classifier and uses the classifier's gradient during denoising steps to guide a frozen π0 policy, enabling transfer to new robot embodiments without additional demonstrations or parameter updates.
The arXiv paper Micro-Localization demonstrates high-precision global localization using microscopic optical observation of natural ground texture, achieving localization accuracy on the order of tens of micrometers. The authors established the relationship between optical magnification and image-space localization uncertainty on the one hand and physical localization accuracy on the other, and addressed the calibration, camera modeling, and focusing challenges introduced by high-magnification imaging.
An arXiv paper proposes RePlanMem-VLA (Replanning-Guided Keyframe Memory for VLA), which turns a policy's own plan revisions into an online memory-write signal: during receding-horizon execution, if consecutive action chunks show unusually large disagreement in their predictions of the same future action, this is treated as a task-relevant event.
The paper proposes RTISSLAM-3D, a 6-DoF SLAM system for a wheeled robot equipped with an in-air 3D imaging sonar, an IMU, and wheel encoders. The system uses a spherical CFAR detector to reduce sonar images to sparse 3D detections, which are used for 3D acoustic odometry that models the sonar beam shape, for place recognition with polar submap descriptors, and for loop-closure verification via 3D registration, all combined into a 6-DoF pose graph with attitude factors.
The paper proposes RAX (Robot Agent eXploration), an open-source robot agent framework centered on scaling online exploration, in which a general-purpose coding agent generates robot control code without using action demonstration data, hand-engineered action primitives, or value-function guidance.
Why it matters: The paper compares the imitation learning paradigm that relies on demonstration data with a demonstration-free online exploration approach, and reports quantitative gaps in simulation and on real robots, helping readers assess the feasibility of this direction.
The SenseNova joint research team released a report evaluating GPT-6-Astra's embodied capabilities in robotic manipulation and navigation. Under a unified protocol, the team evaluated it on 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands, allowing in-context learning (ICL) from training-set demonstrations and substantially increasing the number of execution steps, achieving what the paper calls unprecedented performance.
Mothership proposes a two-tier autonomous refuse collection system in which an electric compaction truck carries two low-speed sidewalk robots, RoboCarts, that fetch and deliver standard two-wheel bins to the truck, return them to their positions, and re-dock for charging. A discrete-event simulation over synthetic campus layouts at 2-16 bins/ha quantified throughput: a two-cart configuration serves 20-29 bins/h, while three to four carts can reach 35-49 bins/h at an 80 m docking radius; the optimal docking radius must be chosen together with fleet size.
SpatialScout is a new benchmark for evaluating vision-language model (VLM) agents' ability to actively build object-level 3D maps. It contains 80 diverse indoor scenes and 3,140 target objects. The task requires agents to control a camera to explore scenes and estimate object centers in a shared coordinate frame, coordinating viewpoint selection, cross-view association, geometric estimation, and map revision within a limited interaction budget.
PhysSkin++ proposes a physics-informed framework that uses a Neural Skinning Fields Autoencoder to learn continuous neural skinning fields from static geometry, constructing a compact deformation subspace without per-object network training or simulation trajectories.
The paper proposes Ropedia W0, an egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. The research builds the Xp-Data dataset, containing 10,000 hours of egocentric human data and 3,000 hours of robot data, with unified multimodal annotations and quality control; camera poses encode locomotion and projected skeletons encode manipulation, with robot skeletons combining each embodiment's forward kinematics and manual keypoint annotations.
ACE-Brain-0.6 combines robot manipulation recordings with complementary visual data from multiple sources through a unified data pipeline; after automatic annotation, the data is converted into verifiable grounding supervision, and Qwen3.5-9B is fine-tuned on about 1.5 million samples to produce the model. It outperforms GPT-6 Astra on four of seven public grounding benchmarks, with an overall average score of 75.41 versus Astra's 73.68. Under a fixed fine-tuning setup, aggregate grounding performance improves steadily as the dataset expands to cover more grounding scenarios.
FocusGUI proposes a lightweight adapter that achieves focused perception for omni-modal GUI agents by dynamically weighting multimodal channels at each step, compressing key evidence into a compact latent, and injecting it into the frozen decoder's embeddings. Across four omni-modal backbone models, it raises Type Match by an average of 33.2 and Exact Match to 3.0x relative to the frozen backbone, and improves Type Match by 30.1 over the strongest existing enhancement method. Training relies only on action supervision, requiring neither modality-relevance labels nor backbone fine-tuning.
The paper proposes Anti-Persistence, a memory design that accumulates trajectories and outcomes, reconstructs task guidance on demand, and discards it after execution, so that errors in stored abstractions are not repeatedly reused. The authors call the phenomenon in which errors are regenerated through a cycle of abstraction, execution, and re-abstraction error reproduction. Across multiple benchmarks and backbone models, Anti-Persistence achieves higher held-out task success rates and the lowest average LLM call cost among the evaluated memory designs.
Researchers propose PVBench, a physical video benchmark that quantitatively measures the physical plausibility of generated videos across 46 dimensions in seven domains, covering instruction and semantic fidelity, behavioral and causal consistency, perceptual and temporal quality, geometric and spatial consistency, material properties and field states, kinematic and spatiotemporal persistence, and mechanics and conservation laws.
UC Davis and ADI have open-sourced Agile-WAM, a tactile world action model for contact-rich manipulation. Through multi-scale, multimodal prediction, it lets vision be supervised over long spans and touch over next-frame spans, achieving a 29.4% relative improvement in overall success rate across five real-robot tasks. On a single RTX 4090, inference latency is 10.35 ms with a control frequency of 96.62 Hz, about 40 times higher than the diffusion policy DP; it achieves best or near-best results on multiple of 9 ManiFeel simulation tasks and Flexiv Rizon 4 real-robot tasks.
A team at HKUST-GZ proposes UniWAM, which uses a Mixture-of-Transformers (MoT) architecture to integrate physical reasoning, future visual generation, and action prediction into one model. Built on three expert modules—Qwen3-VL-2B-Instruct, the Wan2.2-TI2V-5B backbone, and an action predictor—and paired with the EgoANT pipeline for automatic annotation of egocentric human video, it is jointly pretrained on human data, robot demonstrations, and VQA data.
Light-MER is a multimodal emotion recognition compression work by the University of Glasgow, the Institute of Computing Technology of the Chinese Academy of Sciences, and Shandong University. Accepted to the ACM MM 2026 main conference, it compresses an 8B teacher model into an 855M student model, with the training signal drawn from hidden-state geometry rather than answers. The work targets the deployment cost of generative multimodal emotion recognition, with the goal of enabling the model to output paragraphs that explain emotions.
StochasticNCBF uses stochastic neural control barrier functions to achieve safe control of an inverted pendulum, analyzing the stochastic generator, safety control correction, and second-order neural barrier propagation. Reproduction experiments compare the safety performance of three schemes: no correction, deterministic correction, and retaining the stochastic term.
EgoWAM examines a concrete deployment conflict in dual-arm robotic manipulation: first-person human data collected by Project Aria glasses is cheap and abundant, but retargeting human hand trajectories into robot actions for behavior cloning (BC) is difficult, so this work switches the supervision channel from actions to world state.
VLA-Rolling is a lightweight inference framework that reuses the action tails discarded during rolling planning by flow-matching VLA models. Through three stages—action-tail inheritance, per-action confidence estimation, and renoising with denoising step allocation—it reuses denoising progress while keeping the underlying VLA frozen.
PhysLDM proposes a unified latent diffusion paradigm for one-shot high-fidelity deformable body simulation. At its core is a holistic spatiotemporal VAE that avoids the staircase artifacts of standard temporal compression, achieving approximately 2.48 mm reconstruction precision on meter-scale scenes with up to 78x token compression.
The paper proposes BiDex-100, a simulation benchmark for dual-arm dexterous manipulation comprising 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes to assess robustness.
A Perspective article argues that Joint-Embedding Predictive Architectures (JEPA) are moving from representation learning toward latent world models that support prediction, planning, and control, and that interpretability research must shift from "what is encoded" to how latent variables interact, evolve, and causally drive prediction.
AffordanceWAM is an affordance-aware generative world-action model that represents object-centric spatiotemporal affordance through Scalar Affordance and Affordance Heatmap within the generated future world, grounding visual prediction in task-relevant objects and interaction regions.