EMBODIED AI · 2026-11-18

Embodied AI — Paper Daily, 2026-11-18

Today's embodied track asks how a model and its data can cross different bodies, different sensors, and the gap between simulation and hardware. One group of world-action models stops predicting a barely moving background and instead learns what is actually changing in the image. Some revise an imagined future with the actions they just proposed, then tighten the actions. Another group folds heterogeneous touch into a shared representation, sometimes anchored in six-dimensional force and torque, sometimes first mapped into one fingertip coordinate frame. On humanoids, one line turns glasses video into physically verified whole-body trajectories in simulation, without first recording on the robot. Another uses tactile history to open a round doorknob zero-shot on hardware, and one grasping policy aims to span many arms and hands. A survey on the same day warns that a near-perfect score on a simulation benchmark does not mean the model still stands when an object or the camera moves slightly.

4.70

1. Ego2Sim2Humanoid: Agentic Real-to-Sim for Learning Loco-Manipulation VLAs without Robot Demonstrations

Egocentric glasses can record a person's demonstration without occupying a robot, but a person's stature and hand shape differ from the robot's, so copying the hand motion often misses the contact or fails to hold. This method reconstructs the scene from the glasses views and head trajectory, plans stances and grasps the robot can reach without penetrating the object, then completes whole-body motion in a physics simulator and renders training views from the robot camera. Fine-tuned only on this converted data, with no real-robot demonstrations of the target tasks, a whole-body VLA on a Unitree G1 lifts the bottle in nine of twenty trials (45%) and places the hat on the head in five of twenty (25%). Turn a human recording into a physically feasible robot trajectory first, then fine-tune. For teams collecting loco-manipulation data for humanoids, recording can use only glasses and need not occupy the robot.

4.62

2. ResWAM: Learning What Changes, Not What Stays, for World-Action Models

Michigan State UniversityCiscoUniversity of MichiganIBM

Adding a future prediction to an action policy is meant to help the model anticipate how its actions will change the world. Over an action chunk, though, the image often barely moves, so a loss on the complete future is dominated by the static background. This work supervises only the residual from the current frame to a future frame in a frozen visual feature space, and that residual concentrates on the arm and the object being manipulated. The resulting residual world-action model has a few billion parameters and uses no pretraining on robot trajectories beyond the benchmark demonstrations. On real-robot Open, Put and Stack it reaches 80.0% average success, ten percentage points above the strongest baseline, at 92.7 milliseconds per action chunk. Learn what is changing in the image, rather than copying the unmoving background again. For closed-loop use on hardware, compute can be spent on change, without carrying a heavy video generation model.

4.51

3. TMA: Sim-to-Real Forceful Humanoid Manipulation via Tactile-Driven Motor Adaptation

AmazonRice UniversityUniversity of MarylandUT AustinUniversity of Michigan

Humanoid whole-body work in hinged environments has mostly been about lining up geometry. Opening a round doorknob with force also means adapting in real contact, without losing whole-body balance. This method organizes policy learning around contacts the fingers can feel, and adapts motion online from tactile history. An upper-body policy is coordinated with a pretrained walking controller, trained entirely in simulation, and transferred zero-shot to a Unitree G1 with multi-fingered hands. In simulation, door-open success is 80.9% with the full tactile signal and 73.6% with contact presence alone. On four commercial knobs, ten trials each, with no real-world fine-tuning, it clearly outperforms adaptation from proprioception only. On a hand that cannot back-drive force through the joints, proprioception alone does not show how much force to apply. For teams that want a humanoid to open a real door, force and contact history have to enter the adaptation loop.

4.44

4. DexFormer 2.0: Towards a Cross-Embodiment Foundation Policy for Dexterous Grasping

HKUST (Guangzhou)BrainCoChongqing UniversityWuhan University

Dexterous grasping policies are usually trained for one arm paired with one hand, so every new pairing needs its own policy. Robots differ in the number and layout of their joints, and observations and actions share no common interface. This reinforcement learning policy maps every robot to a canonical joint interface with validity masks, biases attention with the kinematic tree, and uses a short memory to infer that body's dynamics from recent interaction. Trained on eighteen arm-hand pairings, it reaches 89.01% stable success on seen combinations and 51.18% on six pairings held out of training. On a physical Flexiv arm, student policies distilled in simulation are deployed zero-shot, with no real-world fine-tuning. Different arms and hands have to be written in one joint interface before a single policy can be reused across bodies. For teams that change hands or arms on a line, this is less work than retraining a policy for every pairing.

4.30

5. Learning Heterogeneous Tactile Representation for General Dexterous Manipulation

Robbyant

Contact-rich dexterous work depends on touch, but piezoresistive, magnetic and visuotactile signals differ in layout, and a policy can hardly ingest them raw. This work first maps each sensor's sensing points into a shared fingertip coordinate frame, trains one encoder with self-supervised prediction in latent space, and produces tactile tokens that plug into existing policies. With the encoder frozen, it raises the average success of four policies by 13.9 to 24.8 percentage points on twenty-six real dual-arm tasks. Adding the raw tactile signal yields at most about half that gain, and sometimes distracts the policy from vision it already uses. Read different sensors into one fingertip representation first, so a policy can use contact instead of being pulled off course by the raw signal. For teams that want to reuse tactile data across hands, the point is to have a representation that already crosses sensors.

4.29

6. TaF-Foundation: Wrench-Grounded Cross-Sensor Tactile Representations

Tactile sensors that work on different principles produce incompatible observations, so a representation is hard to move from one device to another. This work anchors tactile representations in six-dimensional force and torque, pairing every sensor with those measurements rather than pairing tactile signals directly across sensors. More than one hundred million paired observations were collected from twenty-four sensors in four families, and different readings are mapped into one feature space by reconstructing force and torque. On the core benchmark, single-frame force error falls 17.1% relative to the strongest baseline, and cross-sensor top-1 retrieval rises by 13.11 percentage points. Plugged into a policy, average success on four real-world tasks is 46.25%, against 23.75% for the strongest baseline on the same set. To align different tactile sensors, first align the force and torque they share. For teams that mix tactile hardware, this path does not require pairing sensors two at a time.

4.27

7. DiffWAM: Learning World Action Model from Consecutive Visual Difference

World-action models jointly predict a future scene and the robot's actions, but a future state entangles task-relevant change with a large amount of unmoving background, so the model struggles to catch the changes that actually matter for execution. This work predicts the future state and, alongside it, the difference between consecutive observations in latent space, giving temporal change an explicit learning signal. At deployment the branches that predict the future and the difference can be removed, adding no extra inference cost. Average success is 98.8% on LIBERO and 88.3% on three real-world manipulation tasks, fifteen percentage points above the strongest baseline. Learn change as its own training signal, then drop the future-prediction branches at deployment so control does not have to slow down. For teams already using this kind of model but worried about slower inference, this separates the training signal from the path that runs on the robot.

4.20

8. ThinkWAM: Rethinking Latent Futures for Action Refinement

World-action models emit actions while also predicting the future, but a first-pass action chunk is often flawed. If the future prediction is not conditioned on that chunk, it will not reflect how those actions would change visual representations, and it is then hard to revise the actions before they are executed. This work lets the policy first propose a complete action chunk, use it to correct multi-horizon future latents, and inject the corrected futures into a second pass of the same action head, which produces the chunk that is actually executed. Success is 84.76% on LIBERO-Plus without out-of-distribution fine-tuning, and 64.4% on average across three real-world out-of-distribution conditions. Mean two-pass inference is 101.2 milliseconds per action chunk on the real robot. First use the proposed actions to revise the imagined future, then tighten the actions against that revised future. For teams that still want a revision before a chunk is sent, the future prediction should see the chunk about to be executed.

3.95

9. Action Intent as a Bridge: Transferring Manipulation Knowledge from Human-Centric Videos to Robots

Vision-language-action models that generalize manipulation still often depend on expensive real-robot data. Off-the-shelf human videos look like a more scalable path, but existing methods extract latent actions entangled with scene noise, and the transferred knowledge is diluted. This work transfers through an action intent shared by humans and robots. It first extracts intents from video, aligns the two domains in that shared space, and then decodes the intent into executable robot actions. On twelve real-world tasks, average success is 82.08%, above the 70.42% of a baseline pretrained on the same human videos. What should move between humans and robots is a clean action intent, not a latent action tied to one scene. For teams that want human videos to make up for scarce robot demonstrations, agreeing on an intent both sides recognize, then decoding it into joint actions, is more stable than aligning video features directly.

3.49

10. Vision-Language-Action Models: A Life-Cycle Survey of Design, Training and Deployment

Witsense AIUniversity of Huddersfield

Vision-language-action models are already close to a perfect score on a widely used simulation benchmark, but those scores do not mean the model still stands when an object or the camera moves slightly. This survey follows the life cycle of design, training and deployment through eighty-eight works from December 2022 to October 2026. Training data has broadened from teleoperated robot data to community-collected episodes and egocentric human video. Model sizes have split into a mainstream of three to seven billion parameters and a group under one billion that runs on consumer GPUs. The sharper point is that models above 90% on standard LIBERO can fall to zero success once objects move slightly from their training positions. Changing the camera viewpoint or the robot's starting pose cuts success from about 95% to under 30%. A near-perfect simulation score does not show how brittle the model is once the scene changes a little. For teams choosing a model or reporting results, a perturbation suite and clear reporting of training data and latency should sit beside a single simulation total.