EMBODIED AI · 2026-11-06

Embodied AI — Paper Daily, 2026-11-06

Robots need to remember earlier events, choose current motion and determine how much force to exert. WholeBodyMem retains vanished object cues, while Ctrl-F separates force commands from reference motion. RoboARCade adds a reminder: execution still needs independent completion checks. The question is not whether a trajectory looks impressive, but whether observation, decision and execution form a reliable loop.

4.73

1. La-JEPA: Compact Robot Policies with Latent World Modeling

Learning to predict the future during robot training need not require generating it before every action. La-JEPA jointly learns actions and future visual features, but masks future features from the action pathway. At inference it removes future tokens, caches observation KV once per layer and repeatedly denoises only actions. With original-kernel CUDA graphs, the S model’s full policy-call speed still reached 2.6 times optimized Fast-WAM. This measures a policy call, not the complete robot control loop; future supervision primarily shapes observation representations during training.

4.72

2. Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models

Human manipulation videos need consistent action semantics before they can teach robots. Open-HumanWAM aligns hand position, orientation and opening in camera coordinates, filtering invisible hands, unstable trajectories and long idle segments. Under an attention mask allowing action tokens to read the current frame but not future video frames, ten thousand hours of human pretraining followed by the same robot post-training raised randomized-scene success from 2.1% without that pretraining stage to 28.7%. Evaluation selects the best checkpoint separately for each setting; these are not all optimal results from a single model checkpoint.

4.69

3. Perceptive Humanoid Motion from Simulation

Apple

In an unscanned room, finding a chair and sitting stably are different problems. HALO synthesizes vision–language–whole-body motion data in procedurally generated rooms. A high-level controller decomposes tasks, an RGB-based generator proposes motion, and a depth-and-proprioception tracker grounds execution. Single-skill physical evaluation in unseen environments achieved 91% success over 360 trials. This demonstrates transfer for the tested skill repertoire, not the same success rate on long-horizon tasks; motion priors, task-decomposition templates and low-level physical constraints remain important.

4.52

4. Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation

After a robot turns away and returns, identical-looking seats may only be distinguishable through an earlier view. WholeBodyMem uses SAM3 to track task objects, removes camera motion and retains frames of events such as appearance, disappearance or movement. The action model reads those views and object features. Across 60 physical G1 trials, the complete system raised success from the base policy’s 36.7% to 88.3%. The gain includes object features and a decision module, not merely extra frames; cues the tracker never detects cannot automatically enter memory.

4.52

5. Correct Robots Before They Make Mistakes: Proactive Human-in-the-Loop Intervention via Preference Learning

Waiting until a robot errs can fill correction data with unnecessary recovery actions. PHIL overlays predicted trajectories on the physical workspace through an extended-reality (XR) headset so operators can intervene proactively. During takeover the policy keeps proposing motion; rejected predictions pair with subsequent executed corrections from the same state and over the same horizon. Preference training is combined with imitation of original demonstrations. With five correction rollouts per condition, diffusion-policy mean success rose from HG-DAgger’s 44.0% to 70.3%. This combines preview, proactive collection and preference learning; it is not evidence that an interface alone guarantees safety.

4.50

6. A³-VLA: Automatic Perception-Guided Attention Alignment for Robot Manipulation

A robot may understand which object to pick yet allocate attention elsewhere. A3-VLA identifies demonstration regions stable before interaction but changing afterward, then tracks candidate targets with SAM2. Training penalizes insufficient language-to-target and state-to-robot attention mass, while aligning target features with instructions. On RoboTwin, π0’s mean success improved by 9.8 percentage points. The regions provide soft supervision rather than exact ground truth; inference runs only the fine-tuned VLA, without an additional segmentation module.

4.49

7. Ctrl-F: Force as an Interface for Humanoid Loco-Manipulation

Dragging a heavy object requires more than reproducing motion: exerted force must change with load. Ctrl-F uses contact force as the interface between two levels. An object-free simulated tracker learns to resist opposing external forces while following motion; a higher-level planner adjusts force using discrepancies between actual and reference object motion. A G1 hardware demonstration dragged a locked-wheel cart carrying 14 kg of added load but failed at 20 kg. Object state currently depends on AprilTags, and unseen motions require retraining; one force interface does not imply immediate execution of arbitrary motion.

4.44

8. Learning Object-Centric Motion Priors from Human for Robotic Dexterous Manipulation

Teaching a dexterous hand requires checking object motion, not just how human-like the fingers look. This method learns future hand–object trajectories from human interaction data, retargets hand motion to the robot, and uses PPO to reward both robot and object tracking. Experiments cover grasping, articulated-object manipulation and obstacle-aware grasping; ablations identify object tracking as important to learning efficiency. Object models, pose estimates and engineered contact rewards remain necessary. A different robotic hand requires new retargeting and policy training, not zero-training transfer of an identical policy.

4.13

9. Where Is VLA Inference Time Spent? Bottlenecks and Optimization Interactions Across GPUs

Compiling first and pruning tokens afterward does not let their separately measured speedups stack unchanged. VLA-Triage finds that once compilation shortens a stage, later optimizations have different time left to save. For CogACT on A10G, the saving from the same 75% pruning fell from 74.2 ms in eager execution to 33.5 ms after compilation. Quantization can replace kernels, while pruning changes matrix shapes, so combinations need measurement in the deployed execution mode. Timing starts after image preprocessing and ends at action output, not across the full robot control loop.

4.09

10. RoboARCade: Benchmarking Agentic Robot Control

A robot agent declaring completion does not establish task success. RoboARCade separates the agent from privileged simulator state and grading rules, exposing visual, depth or 3D interaction interfaces while an external grader evaluates subtasks. On a subset of tasks with existing successful examples, reading one same-task demonstration raised mean rubric score from 53.5% to 73.0%; this is not full-task success rate. Demonstrations reduce timeouts on bad approaches, but precise placement and false completion claims persist. Explaining a procedure, completing it and recognizing completion need separate evaluation.