EMBODIED AI · 2026-11-12
Embodied AI — Paper Daily, 2026-11-12
Today's main thread in this track is trimming robot policies and checking where their high scores come from. One paper uses training-free nearest-neighbor retrieval to score close to mainstream VLAs on LIBERO, showing that much of the high score comes from looking up demonstrations. APS-VLA keeps only the visual tokens that help actions, and MetaWAM puts all robot signals into a single frame, both cutting latency without changing the backbone architecture. CIALLO and DexSeed find ways to make one set of demonstrations serve more robots and objects. LEASH and NTT's world model focus on the moment before acting: check the scene for safety first, or rehearse the outcome before choosing an action.
4.70
When a video diffusion model is turned into a robot world action model, low-dimensional signals such as joint state, actions and value must enter a latent space that has only seen natural video. The usual fix gives each signal its own frame, which lengthens the sequence and the attention cost.
4.52
A VLA processes hundreds of visual tokens at every step, which makes inference slow, yet the next action often depends on a small part of the image. This paper adds a small scorer in front of a frozen VLA and learns from demonstration trajectories which tokens to keep.
4.52
Robots write actions in different ways, some as joint angles and some as end-effector displacements, so one robot's demonstrations are hard to reuse on another. This paper trains an action codec that aligns actions, the change between two frames and the instruction in one space, which represents the intended motion over the next short stretch of time. Each robot's body differences live only in a small set of its own embeddings, so a new robot needs one recording of at most 10 seconds and about half an hour to update that set. A large model is then trained to predict the intent from images and instructions with only a few demonstrations per task, reaching 76.7% average success on two new real robots, compared with 70.0% for a fully finetuned off-the-shelf robot model. Separating body calibration from task learning lets a new robot get by with very little data. It resembles the recommendation pattern of a shared representation plus a small set of parameters per scenario.
4.49
When a humanoid walks with an object in one hand, the uneven load makes it drift sideways and slowly turn off course, until it may hit obstacles. This paper trains the walking policy with three copies of the environment running together, empty-handed, loaded on the left and loaded on the right, all given the same velocity command. Each copy is penalized by how far its displacement and heading stray from the average of the three. The policy also adjusts its internal features from the estimated external force, while guaranteeing that mirrored observations produce mirrored actions. In simulation, its turning-rate tracking error is more than 80% lower than that of the FALCON baseline. On the real robot, carrying 1.5 kg in one hand during mapless navigation, it reaches the goal without collision in all 40 trials, while FALCON fails every trial through collisions. Comparing the same command under different loads side by side directly suppresses the bias. In recommendation, too, such paired consistency constraints are more direct than simply adding perturbations.
4.42
Dexterous-hand demonstrations are expensive, because teleoperation has to coordinate many finger joints at once. This paper starts from existing human grasp data, groups it by grasp type, transfers each human hand motion to a robot hand, and then varies object shape, start position and placement to expand it into many demonstrations. The key step is training finger control in physics simulation with the robot arm included. A version trained with a floating hand succeeds almost every time on its own, but once the arm's tracking errors come in its success falls below 30%, while putting the arm in the training loop raises it to 95.50%. In the end 19 human sequences yield 25270 demonstrations, and a VLA finetuned only on this data performs all eight grasp types on a real robot when asked in language. Demonstrations used without physical verification teach the policy almost nothing. The same holds when rules or simulators generate training samples: run them in an environment with real constraints first and keep only the successful ones.
4.40
Mainstream VLAs now exceed 95% success on LIBERO, but how much of that is real generalization? The authors build a training-free control: take the current image, find the most similar frames in the demonstration set, and directly execute a weighted average of the actions that followed them, which already reaches 91.7% success. On the perturbed LIBERO-Plus, this retrieval policy loses less than every VLA tested. The authors then use the vision-language backbones of nine public VLAs as retrieval vectors and find that finetuning mainly turns the backbone into a better retrieval vector, while the actions the VLAs output stay close to the demonstrations. On a benchmark with demonstrations this dense, much of the high score can be had by lookup. Recommendation engineers know this well: report a nearest-neighbor baseline next to the complex model, and only then do you know what the model actually learned.
4.39
Dexterous manipulation needs foresight about how the fingers and objects will move next, and many methods get it through explicit future prediction. During training, this paper first learns an action-conditioned world model. It uses scene depth, hand geometry and demonstration trajectories to predict the 3D motion of points on the hand and the change in image features, including fingers the cameras cannot see. It then freezes this model and aligns the policy's internal representations with the world model's predictive ones, transferring the foresight into the policy. At deployment only the policy remains, taking just RGB images, language and robot state, with no depth input and no future prediction. On 11 simulated dexterous tasks, overall success rises from 49.4% to 60.5%. Foresight can be taught to the policy during training, with no need to generate the future at inference. It matches the recommendation practice of training with a teacher that sees future information and serving only the student.
4.18
Letting large models control a robot arm directly produces concrete hazards, such as dropping a phone into a bowl of water or carrying milk over a power strip. The authors build a robot safety benchmark covering five kinds of risk, from instructions that should be refused to injected instructions hidden in the scene. Every task runs in closed loop in simulation and on a real robot, and mainstream models commonly cause harm. Their fix adds three gates around an existing agent: before a task starts, before each action, and before irreversible operations such as opening the gripper or pouring, a fast scoring model judges from the current camera images whether it is safe. The planner and all three gates read the same safety rules. With the gates, the harm rate of one of the best baseline models in simulation drops from 11.23% to 0, and end-to-end time grows by only 1.66% on average. Safety review for robots has to look at the scene, not only the instruction text. Engineers building agent systems can borrow this layered review.
3.40
NNTT
When robots load baggage or stack goods, the main fear is a stack collapsing or an item falling, and existing control programs alone can hardly tell in advance. NTT rehearses before acting: for each candidate action from the control program, a world model generates a video of what will happen next, scores whether the load will collapse or drop, and the most promising action is executed. This world model attaches a physics module to a video generation model, which estimates each object's position, velocity and mass from the video and computes its next motion from the equations of motion. The result guides video generation, so objects rarely vanish or deform and situations missing from the training data can be handled. In a shelf-loading task where a small robot simulates stacking items, success rises from 73.1% to 92.3%. Computing physical quantities explicitly makes the generated future reliable and the basis of the judgment visible. It is the same as scoring candidates in a simulator before launch in recommendation.