Category Frontier Radar

Curated digests of new papers and releases in embodied AI and robot learning, with what each result does and does not show.

Conceptual illustration with no text: a line drawing of a robot arm at the top left; three identical circles side by side, each with a diamond inside and a short horizontal line below, separated by thin vertical rules; two large pale purple circular shapes in the background. No data, axes or interface elements.

Frontier Radar

Frontier Radar | Forward Dynamics Models: future prediction reads action features, actions never read the future, and inference skips future generation

FDM: future prediction reads action features, never the reverse; inference skips the future. Authors report 99.4% on LIBERO, 96% vs π0.5’s 83% on real robots.

October 13, 2026

Conceptual illustration with no text: on a pale lavender background, two identical outlined circles stand side by side, each containing the same diamond outline and each with a short dark rule beneath it; a thin vertical divider separates them, with no arrow between them. At top left is an abstract node-graph glyph, one central circle linked to three small circles, and two large pale circles fill the background. The two identical circles are a neutral symbol for two parallel options.

Frontier Radar

Frontier Radar | RoboAware: learning from same-state counterfactual branches whether modular composition or an end-to-end policy should handle each skill stage

RoboAware learns from paired same-state branches whether modular code or an E2E policy runs each skill stage. Simulation only; authors report 77.0% overall.

October 12, 2026

Text-free conceptual illustration on a pale green background: in the top-left corner, a line-drawn robot arm symbol with three circular joints, a gripper at its tip and a base line beneath it; in the centre, a pale elliptical ring with five evenly spaced outlined circles, each holding a solid dot, around a dark dot on a pale disc; large pale green circles fill the top-right and bottom-left corners.

Frontier Radar

Frontier Radar | LiteNWM: encode once and jointly predict the consequences of 16 candidates and the incumbent trajectory in latent space, instead of rendering an RGB future for each

LiteNWM scores 16 frozen-NoMaD proposals from jointly predicted latents; authors report 17.56% lower error than NoMaD+NWM-XL, 128× faster on an RTX 5090.

October 12, 2026

Warm-toned conceptual illustration with no text: four slanted parallelogram planes are stacked in the centre, each with a short line, suggesting a layered structure from planning to execution; a small fixed-base robot-arm icon is in the top left, and large pale circles fill the top-right and bottom-left corners; the image contains no data, charts or depiction of a real robot.

Frontier Radar

Frontier Radar | Being-M0.7: feeding video-only and motion-only human data into humanoid whole-body control through three training stages and a gated connection

Being-M0.7 pre-trains on 10,000+ h of human video and motion, then adapts to a Unitree G1; authors report 128/180 on SIMPLE and 13/15 real trials.

October 11, 2026

Blue-toned conceptual illustration with no text: several circular nodes placed in order along a path suggest one intervention flow of detecting, retrieving, correcting and handing back control; a generic robot symbol appears in the image.

Frontier Radar

Frontier Radar | RESETTLE: step in only when a frozen policy’s two action samples keep disagreeing, then execute one demonstration-guided correction

When a frozen policy's two action samples keep disagreeing, RESETTLE executes one demo-guided correction; authors report FastWAM 49.3%→58.0% on LIBERO-Plus.

October 11, 2026

Green-toned conceptual illustration with no text: four ringed circular nodes sit on a pale elliptical ring at its top, right, bottom and left; a filled dot on a pale disc marks the centre of the ellipse as a decorative focus, not a fifth stage; a small fixed-base robot-arm icon is in the top left, and two large pale circles fill opposite corners.

Frontier Radar

Frontier Radar | HWAM: a humanoid’s reference action is not the motion it performs, so generate the post-execution body state alongside it

HWAM jointly generates a humanoid’s reference actions and post-execution states; authors report 70.6% vs Fast-WAM’s 43.3% on real-robot Candy Picking.

October 10, 2026

Abstract conceptual illustration with no text: two parallel groups of circles and lines, suggesting two ways of executing that can be compared side by side when the outcome is the same.

Frontier Radar

Frontier Radar | Is Success All You Need?: two VLAs with similar success rates can succeed in very different ways

LIBERO-Spatial, camera perturbation: π0.5 and VLANeXt reach 69.7%/68.0% TSR, yet π0.5’s successful runs show 24.1% more jerk vs 3.2% (authors report).

October 5, 2026

Abstract conceptual illustration with no text: two large pale-blue circles on a light background, four stacked translucent planes in the center, each with a short line mark, and a fixed-base robot-arm icon at the top left.

Frontier Radar

Frontier Radar | ReCo: make a walking policy’s command response predictable first, so MPC can coordinate a quadruped’s base and arm

ReCo trains a walking policy to respond repeatably and gives MPC a fitted response model; authors report 28.7%/27.4% lower RMSE and 75/100 successes in sim.

October 4, 2026

Conceptual illustration: on a light background, a line drawing of a robot arm at top left; in the centre, three outlined circles with dots sit on a pale ellipse, around a larger filled pale circle with a dot at its centre; two large pale background discs at top right and bottom left. There are no arrows.

Frontier Radar

Frontier Radar | Token-World: rolling out a world model in VLM token space to simulate VLA policies without RGB reconstruction

Token-World compresses Qwen3-VL tokens from 2560 to 16 channels and predicts there, no RGB; authors report r=0.794 vs 0.583 for Ctrl-World.

October 3, 2026

Conceptual illustration: on a light background, an abstract robot-arm line linking three joint circles, beside three tilted planes stacked on one another, suggesting the slow planning stream, the fast correction stream and the executed action that combines them.

Frontier Radar

Frontier Radar | DSDyn-VLA: when the target moves, a VLA has to close three gaps in perception, latency and control

DSDyn-VLA pairs a ~2–3 Hz optical-flow planner with ~30 Hz residual RL; authors report 49.0% real-robot success vs 8.4% (π0.5) and 38.1% (DynamicVLA).

October 2, 2026

A mixed-modality input glyph branches to the reasoning and generation towers; the generation tower lists video, audio and action outputs. This is a simplified dual-tower schematic, not proof of AWS deployment performance.

Frontier Radar

Frontier Radar | NVIDIA Cosmos 3 on AWS: reasoning, video and actions in one forward pass, and the validation conditions attached to instance sizing

AWS on Cosmos 3: full Nano/Super checkpoints tested in BF16; H100 8-GPU SFT tested by NVIDIA. The post provides no benchmark numbers.

October 1, 2026

Four-step feedback loop: diagnose a real-robot failure, reconstruct it in simulation, refine the policy, redeploy and return to diagnosis.

Frontier Radar

Frontier Radar | F4R: rebuilding real-robot failures as simulation training grounds, in a loop of diagnosis, reconstruction, refinement and redeployment

F4R diagnoses real failures with a VLM agent, rebuilds them in sim and trains around them; authors report 90.0% OOD success, 18.75 pts over Targeted BC.

October 1, 2026