Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.02968
ReContact: Reconstruct and Retarget Human Contacts from Video for Contact-Rich Humanoid Manipulation
ReContact turns human manipulation video into humanoid motions and policies that preserve when and how support transfers between hands and body surfaces. Its object-centric contact plan records contact phases, robot–object pairs, anchors, pressing directions, and whether each contact sticks or slides. Those constraints guide object reconstruction through occlusion, whole-body retargeting, and residual reinforcement learning over a frozen motion tracker. In simulation, ReContact achieves the intended two-palm bottom hold in 83.5 ± 5.2% of thigh-supported regrasp episodes, compared with 0.16% for the strongest adapted baseline. A policy trained on 247 generated references reaches 97.8% tube-transfer success within its generated workspace.
How it works
The contact plan has phase intervals, active contact pairs, a carrier, and a stable subinterval for training assistance. Each pair binds a robot link and a point and normal on that link to an object-frame anchor and outward normal; its mode is sticking or sliding, with a path for a sliding anchor. Contact keyframes mark support changes, while posture keyframes and foot-support intervals provide additional motion constraints. For exocentric clips, GPT-6 Astra drafts the plan and an operator can review it.
What the paper evaluates
The simulation study covers T1 under-arm box carry, T2 thigh-supported regrasp, T2-R reverse regrasp, T3 tissue-pack under-arm transfer, T4 shuttlecock-tube under-arm transfer, T5 a tube-transfer placement family, and T6a spray-bottle and T6b banana grasps. T1–T3 use exocentric video and T6 uses egocentric video. T5 consists of 247 references generated from T4. The G1 is simulated in mjlab; all compared policies train for 3000 updates of 4096 × 24 transitions. Shared-reference comparisons use three adapted baselines.
Limit to keep in view
The reported quantitative controller comparisons are in simulation; the hardware section demonstrates task execution but does not establish comparable aggregate success rates.
arXiv:2611.09124
Toward the Era of Agentic Embodied Intelligence
A2-MEAS is a multi-agent embodied-control framework that combines the A2-WLA vision-language-action policy with a task coordinator, diagnostic agents, and a learned correction interface. A2-WLA uses six-frame motion-aware visual history, a latent multiview world model, proprioception, and an 18-layer action expert to generate 50-step continuous-action chunks. When execution deviates, MetaBrain formulates a recovery instruction, diagnostic agents provide spatial and interaction evidence, and a correction encoder converts that context, executed history, and a fresh base proposal into a code-conditioned residual. The base policy and correction parameters remain fixed during deployment.
How it works
A2-MEAS organizes execution around a local control loop and a task-level reasoning loop. A2-WLA repeatedly predicts action chunks from the complete task instruction, recent multiview observations, and proprioception. The execution layer dispatches only a prefix, then obtains fresh observations. MetaBrain monitors progress against the full goal and invokes diagnostic functions when an abnormal event occurs. Diagnostic agents cover perceptual grounding, spatiotemporal grounding, and interaction or feasibility assessment.
What the paper evaluates
The main benchmark is BEHAVIOR 2026, a 100-task long-horizon household-control suite built from 20,000 teleoperated demonstrations totaling approximately 1,950 hours across seven scenes. The public evaluation uses ten instances per task, or 1,000 episodes per configuration. The policy receives six frames per view, predicts 50 controls, and executes at most 32 before observing again. The paper reports Q-score, which measures the fraction of goal conditions satisfied, complete-task success rate, and successful task coverage. Higher values are better for all three metrics, but Q-score is not equivalent to completing the entire task.
Limit to keep in view
Correction supervision is mined from unexecuted deviations at demonstration states, whereas deployment failures can change object arrangements, grasp configurations, and contact conditions; this creates a stated distribution gap.
arXiv:2611.07411
Kinematic Hypothesis Context: Embodiment-Conditioned Learned Video Coding for Bandwidth-Limited Robot Policy Uplinks
Kinematic Hypothesis Context (KHC) is an embodiment-aware learned video codec for remote robot policies. Both sender and receiver construct zero-rate visual hypotheses from joint states, robot geometry, and camera calibration, including plane-sweep warps, articulated robot reprojections, an end-effector plane, and background memory. On LIBERO, KHC reduces bit rate by 23.9% versus a capacity-matched history codec at equal PSNR and improves closed-loop OpenVLA-OFT success from 22.8% to 43.1% at 6 kbps; lower-rate training reaches 72.6% at the same budget. The approach adds no learned parameters to its 7.0 M-parameter host, but depends strongly on synchronized state-image timing and usable calibration.
How it works
The shared context contains decoded frames, joint states, prior actions, the robot model, and camera calibration. From the previous and current states, KHC constructs J=D+5 warped references: identity, D inverse-depth plane hypotheses, an articulated robot layer, an end-effector plane, background memory, and an auxiliary slot. The robot layer uses link-specific motion and rendered depth, while the background memory retains regions uncovered by the moving arm.
What the paper evaluates
The main benchmark is LIBERO. Demonstrations are rendered at 256×256 with wrist and static third-person cameras. Training uses LIBERO-90, validation uses the last demonstration of each task, and testing uses 200 episodes from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. The paper also evaluates 100 DROID episodes from Franka robots at 320×180 and 15 Hz, split by laboratory into 63/7/30 episodes; only the wrist camera has usable self-calibrated hand-eye geometry. The coding evaluation uses real bitstreams at stride 8 and reports RGB PSNR, MS-SSIM, LPIPS, and OpenVLA-OFT action deviation.
Limit to keep in view
The strongest evidence comes from LIBERO simulation and a limited DROID evaluation; on DROID, the two exterior cameras lacked usable extrinsics and therefore fell back to the history codec.
arXiv:2611.08156
La-JEPA: Compact Robot Policies with Latent World Modeling
La-JEPA is a compact robot policy that trains World and Action diffusion transformers from scratch in the dense feature space of a pretrained V-JEPA 2.1 encoder. It jointly predicts future visual features and action chunks during training, but removes future tokens at inference so only action generation is repeatedly denoised. Across LIBERO, corrected LIBERO-Plus, RoboTwin 2.0, and four Franka tasks, the paper reports strong or leading performance against larger video-generator and vision-language-action policies, with 0.10–1.14 billion trainable parameters in the LIBERO configurations and 0.74–1.78 billion total parameters at deployment, including the pretrained perception encoder.
How it works
La-JEPA maps multi-camera observations, proprioception, and a language instruction to an action chunk. The pretrained V-JEPA 2.1 ViT-L encoder processes each camera independently. The current observation becomes context features, while features representing a 32-frame future horizon become prediction targets. The paper normalizes and spatially reduces these grids before projecting them into the World DiT embedding space. The visual encoder is the only visual pathway: the policy does not separately consume RGB and does not reconstruct pixels.
What the paper evaluates
The paper evaluates on LIBERO, LIBERO-Plus, and RoboTwin 2.0, then adds latent world-action comparisons, latency measurements, and four physical Franka tasks. LIBERO policies are trained on the four standard suites and evaluated both in-distribution and on LIBERO-Plus variants perturbed along camera viewpoint, lighting, sensor noise, and other axes. RoboTwin policies are trained only on clean demonstrations and evaluated in clean and randomized scenes; the randomized setting adds clutter, distractor objects, and new textures. Success percentage is the primary metric, so higher is better.
Limit to keep in view
The paper does not exhaustively tune learning rates, weight decay, dropout rates, LoRA settings, or the relative capacity of the World and Action DiTs, so the non-monotonic RoboTwin scaling cannot be cleanly attributed to architecture alone.
Read the full reviews
- ReContact: Reconstruct and Retarget Human Contacts from Video for Contact-Rich Humanoid Manipulation · arXiv:2611.02968
- Toward the Era of Agentic Embodied Intelligence · arXiv:2611.09124
- Kinematic Hypothesis Context: Embodiment-Conditioned Learned Video Coding for Bandwidth-Limited Robot Policy Uplinks · arXiv:2611.07411
- La-JEPA: Compact Robot Policies with Latent World Modeling · arXiv:2611.08156
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.