EMBODIED AI · 2026-11-11
Embodied AI — Paper Daily, 2026-11-11
Most of today's embodied AI papers ask the same question: how a general robot model gets from demos to stable deployment. Kimodo-Policy and SWAM change the pretraining base: the first freezes a motion generator as its backbone, and the second pretrains a world action model on large-scale video and human demonstrations and finds that caption detail affects control. DASH takes the slow backbone off the control loop, and CTR retimes bimanual demonstrations to remove false dependencies that policies learn from demonstration order. Tactile sensors can stand in for wrist cameras on contact-rich tasks, and AnchorQ refines frozen policies with a small set of meaningful candidate actions.
4.52
In two-armed robot demonstrations, which hand moves first is often just the operator's habit, yet imitation learning cannot separate the orderings a task requires from the ones that happened by chance, so when one arm is paused the other waits too. The authors audited four public bimanual datasets and found that tasks whose order could change rarely include more than one recorded order, and more data with the same order does not fix this.
4.52
When a humanoid walks while picking things up, its torso, arms and legs must coordinate continuously, yet existing vision-language-action and world action models inherit semantics and visual dynamics from pretraining and still have to learn body coordination downstream. This paper changes the pretraining target: it takes a text-to-motion model for the whole body as a frozen backbone and trains only small branches that feed in camera images, plus hand control. With the same structure randomly initialized and then trained, success on object-interaction tasks falls from 51.67% to 18.33%, which indicates that the inherited motion structure does help. On a real robot the mean success across eight tasks is 60%, while the lead on simulated leaderboards is small. With motion as the pretraining target, downstream training only has to connect the images, the same practice as freezing a pretrained model and training only an adapter layer in recommenders.
4.49
The hard part of making two dexterous hands cooperate is specifying which hand moves and when to pass the object, and past work mostly wrote rewards or skill sequences for each behavior, which had to be redone for each new task. This paper tries another route: the reward is written only around the object, checking whether it reached the goal pose and whether the right hand touches it, and only the training environment changes. For example, when the goal sits in the other hand's workspace, the object cannot get there without a handover, and grasping, in-hand adjustment and handover were all trained this way. Three expert policies are then distilled into one, which reaches 76.3% success on four real-robot tasks with unseen objects and no task-specific tuning. In simulated barcode scanning it reaches 85.0%, while two single-arm policies without handover reach 60.4%. The engineering moves from writing rewards to designing the training distribution, the same trade-off recommender engineers face when tuning sample distributions.
4.49
Wrist cameras are nearly standard for robot-arm policies, yet once the fingers close, the camera cannot see the actual point of contact. Earlier work showed that touch helps but did not answer whether it can directly replace the wrist camera. The authors mount two low-cost tactile arrays on the inner gripper surfaces and feed the readings into the same ACT policy as two extra images, with demonstrations and training settings unchanged. On three contact-rich tasks, retrieving an object, inserting a plug and wiping a whiteboard, the policy with touch and no wrist camera beats the vision-only policy with a wrist camera every time. The difference is significant on two tasks, with 81.7% against 50.0% for retrieval from an opaque bag. Using both inputs does not reliably help further. Vision brings the hand to contact, and the state after contact depends on touch, and when adding ranking features, signals that cover a blind spot are worth the most too.
4.44
Navigation models usually make use of only the last few frames, so returning to a target that left view long ago is nearly impossible, feeding in every past frame hurts attention quality and costs heavy compute, and sampling frames at a fixed interval loses information. Compass 1 draws the path traveled as a 2D map, marks it with keyframe numbers and headings, and feeds it with sparse keyframes as ordinary images, leaving the model architecture unchanged. The map alone does nothing because the backbone ignores it, so training adds question-and-answer supervision that asks which frame shows the target and which direction to go now. Overall success in real-world scenes is 79.4%, against 17.8% for the strongest baseline. On the hardest task of returning to an unseen target, every baseline fails and Compass 1 reaches 58%. History has to be compressed, and supervision has to force the model to use it, which is also the sticking point in long-term behavior sequence modeling.
4.42
World action models predict future video and robot actions together and usually start from video pretraining, and existing work assumes the text-to-image correspondence learned there survives action training, while guidance on how to organize data and annotation is limited. This paper runs item-by-item comparisons on about 47,000 hours of general video, first-person video, handheld gripper demonstrations and robot data, and finds that caption quality directly affects control. With the same training budget, rewriting subtask captions from one verb-object phrase into a step-by-step account and adding scene and lighting details raises success from 47.20% to 55.04% on simulated tasks with changed backgrounds and lighting. Training on a mix of first-person video and handheld gripper demonstrations also beats using either one alone. More detailed captions end up as better actions, so anyone doing multimodal pretraining should audit annotations early.
4.37
The backbone of a vision-language-action model is slow and the robot does not stop to wait, so actions are computed from stale images, the problem grows with backbone size, and deployments settle for small backbones. DASH splits the slow backbone and the fast action head into concurrent processes: the action head reads the backbone's latest semantic output and combines it with fresh image features from its own vision encoder, so the action head alone sets the control rate. From the smallest backbone to the largest, the control rate holds at 20 Hz, while the synchronous setup falls from 9.7 Hz to 4.1 Hz. Splitting alone is not enough: the asynchronous version trained on aligned data is no better than synchronous, and success exceeds synchronous only when training deliberately pairs the action head with old semantic features and new images. Time offsets in training samples must match production, just as training and serving feature freshness must match in recommenders.
4.20
A common way to improve a frozen general robot policy on a specific task is to use reinforcement learning to search the initial noise it generates actions from. But directions in noise space have no regular link to actions and online interaction is costly, so exploration is inefficient. AnchorQ first maps coarse motions such as moves in six directions back into noise, which gives a small set of meaningful candidates at each step (the paper calls them anchors), then samples refinements near them, and a value network picks the highest-scoring one. With the same 500,000 interaction steps, a simulated task suite reaches 95.5% success, against 48.3% for DSRL, which searches the noise directly. On a real robot with a small interaction budget, it also beats DSRL on all three tasks. Marking meaningful candidates first and then refining locally reads like the retrieve-then-rerank pattern in search, and it saves far more than blind search in a continuous space when the exploration budget is tight.