4.55
Teaching a humanoid a long task such as walking to a table, picking up a banana and putting it in a basket usually takes a hundred or more teleoperated real-robot demonstrations, and moving the table often means collecting them again. Zero-HLM builds all of its data in simulation instead. It scans a real room with a smartphone, rebuilds it as a simulated scene that looks like a photo and supports physical interaction, and rewrites a few teleoperated demonstrations into a thousand or more new ones. Each new demonstration randomly moves the robot's starting point and the positions of the object and placement target, the hand trajectory moves along with the target, and only those that succeed in simulation are kept. A policy trained only on this data and deployed on the real robot averages 66.7% success over three tasks, above the 46.7% from training on real demonstrations. At positions the real demonstrations never covered, success rises from 23% to 69%.
4.51
To learn to push objects or turn an object in hand to any pose, a dexterous hand usually needs either demonstrations for each task or reinforcement learning in simulation followed by transfer to the real robot. PlayPolicy, from Amazon FAR and collaborators, has a person teleoperate the hand to fiddle freely with objects for a few minutes, and uses behavior cloning to learn a play policy tied to no task. The robot then plays on its own for hours, and this data trains a world model that predicts what actions will do. For a specific goal, the play policy proposes candidate actions, the world model predicts each outcome, and the one closest to the goal is executed. On the real robot its success rate is 69.6% higher than behavior cloning with goals relabeled afterwards, and it matches reinforcement learning with about 490 times less interaction data.
4.50
Video diffusion models have seen a great deal of how objects move and get pushed, but using that understanding for robot control first requires solving two things. Datasets differ in the number and placement of cameras, and robots differ in their action formats. X-World, from Fudan University, adds a cross-view attention branch to a pretrained video model so it can watch one to three cameras at once. It then compresses the video features into a fixed-length sequence of tokens for a shared action module, and each dataset gets a lightweight projection layer that converts the shared module's output into that robot's own action format. It averages 98.5% success on LIBERO and reaches 81.7% on six real dual-arm tasks, above the 58.3% of the strongest baseline, X-VLA. Removing video adaptation or cross-view fusion causes the largest drops in success, which shows that the video prior is what does the work.
4.47
Robot manipulation often fails in the last few millimeters. The policy recognizes the object to grasp but misses the pose offset, the object's edge, or the position just before contact. The paper argues the problem lies in the visual encoder, because common encoders aim for semantic invariance and wash out these small geometric differences. Diffusion model features are highly sensitive to pixel-level correspondence, but they are noisy, slow, and unusable as is for real-time control. Robot-DIFT distills a diffusion model that takes noisy input into a deterministic network that takes a clean image and produces features in one forward pass, then fuses its coarse-to-fine features into visual tokens for the policy. On RoboCasa, LIBERO-10 and real robots, it beats vision-language, self-supervised, geometry-oriented and diffusion encoders on tasks that need precise contact.
4.46
Pushing a cart, moving a heavy box and wiping a whiteboard require controlling not just how the hands move but how hard they press. On a humanoid this force often has to come from both hands and the whole body together, and motion trajectories alone cannot describe it. FLoRe collects demonstrations with two force-sensing handheld grippers and wearable trackers, recording hand motion, contact force and whole-body motion without any robot. Besides predicting actions, the policy outputs a force command for each hand. While keeping balance, the whole-body controller makes each hand hold that force along the commanded direction and stay compliant in the others. On three contact tasks on a Unitree humanoid it succeeds 80% of the time, while a position-only policy trained on the same demonstrations reaches only 26.7%. For humanoids that must push, wipe and carry, force belongs in both the demonstrations and the policy output.
4.40
To give a VLA model, which sees images, reads instructions and outputs actions directly, a sense of 3D space, the common approach is to align its internal representations directly with features from a pretrained 3D foundation model. The paper finds that this point-wise alignment imposes a low-rank constraint on the feature space, causing severe rank collapse in the VLA's original representations and losing their expressive power. RIGID, from Queen Mary University of London and collaborators, aligns only relationships instead. However similar the features of any two positions are in the 3D model, the VLA's features at those two positions must be equally similar. Where each feature points and how long it is no longer matter, so the VLA keeps its own feature coordinate system. On public benchmarks, a new spatial reasoning benchmark built by the authors, and real-robot manipulation, it brings 3D spatial information into VLA reasoning and generalizes more robustly. Teams adding 3D priors to a VLA should first check whether their representations collapse after alignment.
4.39
RoPE in robot policies usually computes relative positions from a token's index in the sequence or grid. Yet the same pixel offset maps to different viewing angles in different cameras, and the same frame offset can mean different elapsed times. An image token also covers a patch of pixels, and a compressed video token mixes several moments, so a single center point describes it poorly. POD-RoPE takes several observation points within the pixel region and time span each token covers. Each point's camera position, viewing direction and sampling time form one coordinate. Just as ordinary RoPE multiplies an index by frequencies, the coordinate is multiplied by a set of fixed frequencies to get rotation angles, and the averaged cosines and sines of these angles over all the points become the token's positional encoding. Added to Fast-WAM, it raises mean success on two real-robot tasks from 32.5% to 52.5%, with only about 17% more latency than ordinary RoPE.
4.38
Sampling several candidate action chunks from a VLA and letting a scoring model pick the best is a way to raise success without more data. Common scoring models, however, learn from whole-trajectory success or failure and treat every moment the same. During free-space motion or carrying after a grasp, the sampled candidates are nearly identical and any choice works. Only at a few key moments, such as aligning before a grasp or a placement, do candidates differ enough that a wrong pick causes failure. At each moment DiVeR samples several candidate actions from the VLA, takes for each candidate the average of the action tokens' vectors in the action module's last layer at the final denoising step, and averages the per-dimension variance of these vectors. The higher the variance, the more critical the moment. Training the scoring model then gives critical moments more weight by this score, with no per-step labels. On four real tasks with a Franka arm, mean success rises from 58.3% with a single sample to 71.9%.
4.37
Thin deformable objects such as ropes and cables are hard to manipulate for two reasons. They are so thin that vision struggles to locate them precisely, and how they deform depends on physical properties such as stiffness and thickness. PS-DLO first has the policy judge the rope's properties from the initial image, such as how many ropes there are and how stiff they are, as a prior for generating actions. During training, a teacher network looks only at the rope region SAM3 segments in the next frame and encodes it as a set of features. The policy sees only the current frame, but its internal representation of the rope must stay close to the teacher's features, so it learns to anticipate where the rope will be next and what shape it will take. On nine real tasks on single-arm and dual-arm platforms, including untying several kinds of knots, straightening, wrapping and threading, it averages 68.9% success, 20 percentage points above the strongest baseline. When handling cables and ropes, giving the policy object properties and future shapes explicitly is a path worth trying.
3.65
GoogleGoogle DeepMind has brought Gemini Robotics ER 2, its high-level reasoning model for robots, to Google Cloud in public preview. The model acts as the robot's brain. It understands space and task progress in video, breaks down long tasks that run for minutes and involve hundreds of decisions, and calls lower-level VLA models and hardware APIs to carry them out, while accepting an operator's change of instructions mid-conversation. On the cloud, companies can test in simulation with MuJoCo or Isaac Sim on the same platform, use the model's pointing, moment finding and progress judgment to label demonstration videos, and then train their own VLA. For teams deploying robots, high-level planning can now come from a hosted model while they focus on low-level actions and hardware.