Embodied AI Digest

• GEN-1.5 leverages natural repetitions in human data (symmetry, recovery) for organic policy improvement
• UMI captures direct human gripper...

A unified reinforcement learning-based controller enables humanoid robots to learn vision-driven reactive soccer skills. This marks a concrete step toward real-time, reactive control in dynamic settings.

Two perspectives clash on humanoid progress:

• Independent benchmarks are essential to validate claims and speed adoption.
• Data scarcity is the true...

China is deploying capital into world models far earlier and at larger scale than during the 2023 LLM wave.

• Funding contrast: At least $1.44...

HarnessEval-W brings the harness paradigm to world model evaluation, deploying specialized sub-agents that generate transparent, auditable reasoning chains for every score. This fills a key gap in rigorous benchmarking.

Robot foundation models have split into Vision-Language-Action models (VLAs like OpenVLA, pi0) that inherit semantic knowledge from internet VLMs and...

PRM-as-a-Judge 1.5 moves embodied AI assessment beyond binary success rates by converting rollout videos into dense progress curves and multiple fine...

Two startups are accelerating robot learning by capturing human demonstrations at scale, moving beyond simulation and teleoperation.

• Generalist...

China's humanoid robot output is projected to hit the 100,000–200,000 unit mark in 2026, yet the "Year of Delivery" is forcing a reality check on whether mass deployment timelines can hold.

Only three of 23 humanoid teams completed the full Beijing firefighting challenge within 30 minutes.

• Sequence tested: Identify random hazards,...

Kinematic errors in humanoid tracking miss the contact and stability failures humans notice most, prompting the new HumanTracker benchmark with 153 hours of diverse motion data and a preference-trained HumanScore that better matches perception.

• LG's bipedal humanoid, built on NVIDIA Isaac GR00T and Jetson Thor, targets Q1 2027 public reveal while entering factory testing.
• CLOiD wheeled...

Dyna-2's dual next-frame/next-action design, trained on 1M+ hours of human video, delivers predictable gains with no plateau from 1k–1M hours and...

Fi0 addresses robot AI's hardware dependency by representing each robot's morphology and state as input for action generation, allowing task knowledge...

A clear trend is consolidating around world models that learn physics and actions directly from video, moving beyond vision-language-action...

DexTeleop has become the first company to achieve large-scale embodied AI deployment in open real-world environments through its strategic partnership...

Spatial Memory Agent lets frozen VLMs distill verified spatial experience into reusable lessons with Transfer Reliability Scores, boosting performance across five benchmarks without parameter updates or external tools.

The same handful of robot demonstration metrics—spectral smoothness, jerk, idle-timestep detection—keeps getting reimplemented across unrelated labs...

The open-closed divide in 2026 world models shapes physical AI more than raw capability.

• LTX-2.5 is the sole open-weights model running in...

X Square Robot's WALL-B embodied AI model and 6-axis arm autonomously sorted variable parcel piles at 1,816 parcels per hour with over 98% accuracy in a live stream, adapting to changing conditions where lab demos typically fail.