EMBODIED AI · 2026-11-10

Embodied AI — Paper Daily, 2026-11-10

Today's embodied track keeps the already published manipulation and control cards and adds newly surfaced work. The new cards cover progress-informed retrieval of demonstrations for in-context adaptation, inference-time guidance that steers a diffusion policy around unseen obstacles, and whole-body control that separates tracking torques from a learned balance residual.

4.44

1. Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation

General desktop manipulation has to learn a behavior prior from messy data, transfer across robot bodies, recover after a closed-loop mistake, and keep task state for longer. This work turns large vision-language collections and robot trajectories that cover twelve embodiments into executable control in three stages. It first pretrains the backbone with discrete action tokens. It then replaces the discrete head with a continuous action head initialized from scratch and uses flow matching on unified end-effector action chunks for cross-embodiment control. Finally it post-trains on the real robot with successful demonstrations, corrections in which a person first decides whether to take over and then supplies the right motion, and autonomous rollouts labeled by outcome. As a standalone policy it reaches 93.6% success on RoboTwin 2.0. Data collection rolls back the segment where the policy failed and lets a person attach the correct motion; after training on those corrections, Ethernet-insertion success reaches 70%.

4.44

2. On Agentic Manipulation Policies

Robot manipulation does not have to rely on specialist policies trained on large sets of demonstrations. In this paper a generalist AI model acts as an agent, commands the robot arm directly and adjusts as the task unfolds, with no demonstrations and no task-specific training, and it succeeds far more often than specially trained VLA policies.

4.42

3. VICTR: Value-Informed In-Context Learning for Vision-Language-Action Models

Robot foundation models can adapt from demonstrations placed in context, without further fine-tuning. Retrieval that relies on visual similarity alone may not match the current stage of execution. This method estimates task-relative progress from a reference demonstration video, a language instruction, and the current observation, then uses that progress to retrieve demonstration context. Conditioning on a reference video lets it align the current state with a demonstrated execution even for tasks unseen in training. On real-world manipulation tasks, success rises from 6.2% for a fine-tuned baseline to 23.8%.

4.42

4. SURE:Spatiotemporal Uncertainty Representation for Efficient Vision-Language-Action Learning

Vision-language-action policies are hungry for demonstration data. In manipulation, several nearby motions can all succeed, and similar demonstrations often unfold at different speeds, so treating every demonstrated frame as the only correct target makes the model learn unnecessary detail. This work discretizes each translation and rotation dimension into many ordered bins and uses a shared-weight transposed convolution to score the bins so neighboring bins get related scores. Those scores are turned into probabilities, and the continuous action is the probability-weighted average of the bin centers. On LIBERO-Spatial, success with one-tenth of the demonstrations is 21.4 points above OpenVLA-OFT using the same backbone; with one-quarter of the data, success reaches 95.0%.

4.41

5. ForPC: Force-Aware Whole-Body Control for Precise Humanoid Cart Pushing

If a humanoid can push an existing wheeled cart, it can move heavy loads without rebuilding the workspace. Once both hands grip the handle, the same contact forces that drive the cart also disturb the robot's balance, and a change in payload changes that coupling. This work learns force-aware whole-body control with two-stage reinforcement learning. The first stage learns a locomotion prior under varied hand loads and upper-body poses, then freezes it so the later task policy can treat that prior's outputs as already-learned support under load and only learn bounded residual adjustments. During second-stage training, a teacher sees privileged contact forces and motion states in simulation and encodes them as latents; at deployment a student infers those latents from onboard signals such as joint motion, torques, and cart heading. A residual policy then coordinates only bounded adjustments of the arms, waist, and legs together with base speed. Versus a walking-only baseline, cart-velocity error falls by as much as 68.2%. Trained only in simulation, the policy transfers zero-shot to a real Unitree G1 and pushes a cart loaded to 30 kilograms.

4.39

6. Horizon-Coupled Obstacle Guidance for Relative-Action Diffusion Policies

After deployment, a robot often has to avoid obstacles it never saw in training, without collecting new demonstrations or retraining. Existing policy-level guidance often corrects each horizon step on its own, which can fix a local violation without assembling a complete, task-directed detour. This inference-time method guides a pretrained relative-action diffusion policy by rolling cumulative relative actions into a workspace trajectory, scoring obstacle interactions along that path with a signed-distance model, and sending each obstacle response back to all contributing action increments during denoising. In seven obstacle-constrained scenes across four manipulation tasks, it completes 75.0% of trials without collision, compared with 5.7% for absolute-action point-wise guidance.

4.37

7. STEP: Spatiotemporal Training with Explicit Prediction for Perceptive Humanoid Locomotion

A humanoid climbing stairs or crossing gaps reads the terrain through a depth camera, but which part of the terrain matters depends on the robot's own body state and on how both change after each action. Existing methods either fuse visual and body-sensing features or predict one signal from the other at the same moment, and neither directly constrains how the state changes after an action. STEP adds two prediction tasks during training. It hides one signal and predicts its features from the other, and it predicts the next fused state from the current state and the executed action. On a real robot it succeeds in 98 of 100 trials across stairs, a high platform, a wide gap, stepping stones and walking backward on stairs, where it cannot see the terrain behind it. When camera input drops out or lags, both prediction tasks also keep the policy steadier.

4.30

8. ODEAction: From Predicting Chunks to Evolving Actions

Vision-language-action models usually predict actions in chunks, so continuous execution is cut into separately generated segments that can jump where one chunk ends and the next begins. This work stops predicting chunks and instead learns a transition field over action space: the transition vector is the rate at which the current action changes with integration time. At execution the current executed action is the initial value, and a numerical integrator such as forward Euler steps along this field. Each integration step yields the next action to execute, which then becomes the start of the following step, instead of predicting a whole future chunk at once. The method is an action head that is not tied to one vision-language backbone. On LIBERO, EBench, RoboTwin, and real robots it beats matched chunking baselines, raising average success on EBench by 61.8% relative.

4.29

9. DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Control

Humanoid whole-body control has to track diverse reference motions while staying balanced under gravity, torque limits, contact switches, and disturbances. Reinforcement-learning trackers often let one policy output both tracking and balance commands, even though similar reference poses can demand very different dynamics, which complicates learning and can produce high-frequency fluctuations. This framework splits the job. Inverse-dynamics feedforward and joint proportional-derivative feedback supply the tracking torques, and a residual policy learns only the extra joint-torque corrections needed for balanced execution. In simulation and on hardware, one learned balancing policy generalizes across motion datasets and keeps tracking and balance under attached payloads and physical disturbances. Across all evaluated training-data scales it achieves higher tracking accuracy and success rates than a retrained SONIC baseline.

4.14

10. Spatial-UMI-Dream: Scaling World Action Models with Metric Geometry for UMI Manipulation

Collecting demonstrations with a handheld gripper, the UMI approach, is cheap and independent of any particular robot, but the gripper camera is close to the objects and moves with the hand. The image cannot tell camera motion from object motion, so the real distances and sizes that precise manipulation needs must be guessed indirectly from RGB. Spatial-UMI-Dream pretrains a world action model on UMI data with depth, jointly predicting future images, metric depth and actions. The authors also build a simulation benchmark that tests changes of robot and environment. When both the robot and the environment change, it reaches 72.15% success, 10.30 percentage points above the strongest baseline, and it averages 85.0% success on six real tasks. Adding depth makes handheld demonstrations transfer more easily to different robots.

3.95

11. Should the Action Expert Train the VLM? Action-Token Supervision and Gradient Routing in Mixture-of-Transformer VLAs

Popular VLA training recipes block the gradients from the action expert to the vision-language model, and this paper finds that this default is not always right. Such models pair a pretrained vision-language model with an action expert that outputs continuous actions. Blocking gradients is meant to protect the vision-language model's existing abilities, and the recipes also train it to predict discrete tokens encoding the same actions. The authors vary these two choices separately and find that blocking gradients looks harmless mainly because the action-token supervision fills the gap. In open-loop driving, blocking or not makes little difference when that supervision is present. Without it, letting gradients through cuts trajectory prediction error by about 25%. When both parts are trained together from the start, letting gradients through raises closed-loop driving route progress by about 17%, or about 40% without action-token supervision, and robot manipulation improves as well. Blocking gradients should not be the default recipe.

3.95

12. From Pixels to Actions -- Towards Native Vision-Language-Action Policies

VLA models usually compress an image into features with a vision encoder before reasoning about actions, and this paper finds that this step loses information control needs, so removing the separate encoder gives better and steadier control. Manipulation often depends on exactly the details most easily lost in compression, such as local geometry, edges, depth and subtle state changes. NEO-VLA removes the separate vision encoder. Images pass through light convolutional patch layers and then travel with the language through the whole backbone, so the action expert can draw instruction-aware visual representations from different depths. In controlled comparisons, this native design consistently beats encoder-based versions, and neither full fine-tuning nor scale alone closes the gap. A control policy should not see only the vision encoder's final output.

3.87

13. WorldContact: A Contact-Centric World Model for Scalable Robot Learning

Adapting robots to new objects and tasks takes a lot of interaction data, and physics simulation of deformable objects such as shopping bags is slow because the simulator needs tiny time steps to resolve fast motion and keep objects from passing through each other. WorldContact trains a contact-centric world model from a small set of high-quality trajectories and predicts object motion with larger time steps, rolling out states 10 times faster than the source simulator, excluding rendering and disk time. After an existing VLA policy is fine-tuned on data expanded with it and deployed directly on a real robot, single-attempt success at lifting a bag rises from 65% with simulation data alone to 95%. For deformable-object manipulation, a learned world model can be a cheap source of data.

3.87

14. SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

VLA models mostly inherit their visual representations from 2D vision-language pretraining, and spatial understanding is usually added only during task-specific fine-tuning, so it must be redone for each task. SLA moves spatial learning into continual pretraining. Starting from a pretrained π0.5 policy, it learns to predict actions while aligning intermediate visual features with spatial representations from a frozen 3D perception model. The authors also unify robot data from many sources, and during training they randomly drop the robot-state input so the model relies more on images. These extra parts are used only in training, and the inference architecture is unchanged. Under eight real-world deployment conditions, it has the highest success rate among the compared methods in every one.

3.57

15. Boston Dynamics adds wheels to Atlas humanoid to test whole-body control

Boston Dynamics has put wheels on the feet of its Atlas humanoid as a testbed for whole-body control, and it says plainly that this is not a product. The only new hardware is a custom adapter that mounts an off-the-shelf scooter rim on the existing ankle actuator, and in software each wheel is simply modeled as an ankle that can spin continuously. Classical model-based motor control keeps the robot stable, while a whole-body reinforcement learning policy adjusts pose and force in real time from body sensing and velocity commands to handle unmodeled disturbances and poor contact. All behaviors were learned in simulation and run directly on the real robot, so Atlas can roll fast on flat ground and switch back to stepping over obstacles. The engineer's takeaway is that combining the two kinds of control is more agile than either alone.

3.35

16. Make a Robot Arm Feel Weightless with NVIDIA Holoscan 5

When you guide a robot arm by hand and its motors supply just enough torque to hold up its own weight, you can move it with little effort. This is gravity compensation, the starting point for learning from demonstration and working alongside people. It requires computing inverse dynamics from the current joint state continuously and with low jitter. This NVIDIA tutorial builds a three-node control graph in Holoscan that reads joint state at a high rate, computes inverse dynamics with Pinocchio and sends torque commands to the arm. The key is to run the control nodes on one thread to avoid cross-thread handoffs, with queues that keep only the newest sample. Pinning the control thread to one CPU core then cuts the latency of the slowest few runs in ten thousand to about one fifth in the C++ version and by about 9.6 times in the Python version. Jitter in real-time control often comes from runtime scheduling rather than from the algorithm itself.

3.25

17. AI WAVE Wrap-Up: AgiBot Reports Results of Its 24-Hour Convenience-Store Livestream, Moving Physical AI Into Real Settings

AgiBot

Whether robots can keep working in a real store that is open for business is harder to prove than whether they do a single task well. AgiBot kept 6 Genie G2 robots on duty in a convenience store for a full day under a live stream, completing 1,730 tasks in total, and 4 of them passed the assessment and stayed on. The stream closed AgiBot's AI WAVE series of releases, which also included a new dataset release, a simulation platform, an embodied agent framework and a foundation model among seven results. The public account gives only the total number of tasks and the number of robots kept, without success rates by task type, but long, continuous tests in real settings like this are a step embodied AI must pass to move from demos to real jobs.