Agents · industry

Agents in Robotics

Agent deployments and launches in Robotics, from the Drops.

From the Drops

AgentsRoboticsTestingHuman in loop: no

Preprint: a self-editing harness lifts a robot planner to 51.38% on RoboMME

Researchers report in a preprint that SERA, a VLM planner with an editable harness that tracks task context and revises itself from rollouts, reaches 51.38% success on RoboMME versus 44.38% without self-evolution.

Why it matters: It suggests harness changes, not a new model, can help agents on tasks that depend on history. The result is one benchmark with a fixed VLA executor.

Hugging Face Papers · Oct 21

From Drop #046 · Sun, Oct 25 · Evening

AgentsRoboticsTestingHuman in loop: unclear

Preprint: in TARSphere's 360 robot episodes, general models can beat embodied ones

Researchers report TARSphere, a robot simulator, harness and 360-episode benchmark in 30 indoor scenes. Testing 8 multimodal models, they find general-purpose ones can beat embodied models; task updates and human interaction stay hard.

Why it matters: Frequent collisions and falls led the authors to argue robot agents should be scored on execution safety, not just task success.

Hugging Face Papers · Oct 21

From Drop #046 · Sun, Oct 25 · Evening

AgentsRoboticsTestingHuman in loop: no

Preprint: post-training lifts a 9B robot planner from 18.1% to 71.0% success

Researchers report AgenticRoboRL, which pairs fine-tuning with multi-turn RL across robot harnesses, raised a 9B planner to 71.0% on LIBERO-PRO without external memory, near the 72.1% Codex (GPT-5.5) reached with it.

Why it matters: The recipe that improved coding agents may carry over to robot planners, the authors suggest, without leaning on closed models. They plan to release checkpoints.

Hugging Face Papers · Oct 20

From Drop #045 · Sun, Oct 25 · Morning

AgentsRoboticsTestingHuman in loop: no

Preprint: the best VLM agent scores 41.71% at mapping 3D scenes by exploring

Researchers built SpatialScout, 80 indoor scenes with 3,140 target objects. The strongest model tested, Claude Opus 5.5, scored a mean overlap of 41.71%, and only 3.25% of targets cleared an IoU of 0.9.

Why it matters: Agents can recover rough layouts but not precise maps, the authors say: a limit to know before trusting agents with spatial tasks.

Hugging Face Papers · Oct 19

From Drop #045 · Sun, Oct 25 · Morning

AgentsRoboticsTestingHuman in loop: unclear

Preprint: a coding agent driving a two-arm robot sets a RoboDojo record at 39.7%

Researchers report HexaShell, a thin harness where Codex with GPT-6-Astra writes and fixes its own tools after failures, reached 39.7% success across all 42 RoboDojo tasks, a new state of the art.

Why it matters: The limit was contact, not reasoning, the authors say: pushing hit 92.5% but tight insertion only 8.3%. The agent usually saw its failures but couldn't make the motion work.

Hugging Face Papers · Oct 16

From Drop #035 · Tue, Oct 20 · Morning

AgentsRoboticsLiveHuman in loop: unclear

Nvidia extends its Halos safety stack to humanoid and warehouse robots, Ars says

Ars reports Nvidia's Halos for Robotics pairs a dedicated safety processor, monitoring software, sensor checks and simulation. Agility Robotics was first to build it into its Digit 5 humanoid; Boston Dynamics is an early partner.

Why it matters: Nvidia's robotics lead told Ars that safety is the next bottleneck for robots working beside people in factories and warehouses.

Ars Technica · Oct 8

From Drop #012 · Thu, Oct 8 · Evening