Robot progress becomes inspectable from policy to power

Astra mastered placement but stalled at precision

Robocurve ran GPT-6 Astra and two Claude models through 120 trials on the same dual-arm policy and two tasks. Astra placed a block in a bowl in 19 of 20 runs, yet completed only two of 20 precision puzzle insertions. Every run includes a transcript, video and replay artifact.

The 95-percent result is real, and so is the 10-percent result. The gap is the story.

Worth doing: Use the public failures to separate perception and approach quality from the final force-sensitive insertion phase in your own evaluations.

Keep in mind: The study covers two tabletop tasks, used noninterleaved runs and relied on operator-visible human grading.

Facet-0 predicts the force its action will cause

PINE Lab released the Facet-0 model and ManuFacet-1K dataset for tight-tolerance computer assembly. The policy pairs proposed action chunks with the future wrist-wrench trajectory they should create, then uses force-conditioned adaptation during contact. The model, synchronized dataset, paper and evaluation details are public.

Contact-rich control becomes more inspectable when the policy must predict the force signature of its own next move.

Worth doing: Inspect sensor synchronization, task coverage and force ablations before adapting the method to a different wrist or assembly tolerance.

Keep in mind: The released results are limited to the authors’ hardware, tasks and data distribution and have not yet been independently reproduced.

EngineAI opens its humanoid navigation stack

EngineAI released an end-to-end PM01 navigation stack that turns ordered LiDAR range images into velocity commands. The repository covers PPO training in Isaac Lab, ONNX export, Gazebo simulation, separate ROS 2 workspaces and a real PM01 plus Livox launch path. It also documents dependency traps that determine whether deployment works.

The practical value is the full path across toolchains, including the mundane workspace boundaries where robot software often fails.

Worth doing: Reproduce the simulation workspace first, then validate ONNX inputs and LiDAR ordering before enabling commands on physical hardware.

Keep in mind: The repository does not provide an independent field benchmark or broad obstacle-distribution evaluation.

Palmimo gives MCP agents a physical body

Jizai opened the software stack for Palmimo, an assembled six-legged tabletop robot. The repository includes Python drivers, a compute-only dry-run motion engine, MCP tools, agent examples, diagnostics and LeRobot plugins. The software-to-servo boundary is inspectable even though the mechanical design files remain closed.

A useful devkit exposes the awkward boundary between an agent request and a safe motor command.

Worth doing: Use dry-run mode to log every tool-to-motion transition before connecting an autonomous agent to live servos.

Keep in mind: The hardware is not open source, and the repository does not establish autonomous reliability outside the supplied examples.

TRON 2 documents autonomous charging’s last meter

LimX published the first autonomous charging dock manual for TRON 2. It specifies closed-loop return, alignment, charging and departure; a 200-to-650-millimeter capture range; average docking within 30 seconds; dark-scene support; multi-dock identity; SDK calls and electrical protections. These are the details unattended operation depends on.

The last meter to a charger is unglamorous infrastructure, which is exactly why documenting it matters.

Worth doing: Test capture geometry, lighting extremes, dock identification and failed-contact recovery before scheduling truly unattended missions.

Keep in mind: The timing and operating claims are manufacturer specifications, not independent fleet measurements.

Figure commits billions to future robot training

Figure and Nscale signed a multi-year agreement with an initial $3.5 billion compute commitment, an intention to exceed $6 billion and potential deployment of up to 100,000 Vera Rubin GPUs from the second half of 2027. The declared pipeline joins Helix training, Isaac Sim validation and onboard NVIDIA deployment.

This is fallback coverage because the concrete agreement matters, but future compute must not be mistaken for present capability.

Worth doing: Treat the numbers as capacity planning signals and separate signed initial commitments from later intended deployments in roadmap assumptions.

Keep in mind: The largest GPU count and later spending are forward-looking, and the announcement does not demonstrate improved robot performance.

No specialist filler today.

Nothing beyond the general briefing cleared the bar for this role.