Optimus line, VFA funding, lab agents

A production line, a funding round, a purchase order and a benchmark moved this week, each answering a different question about manipulation.

Tesla starts Optimus production in Fremont

Tesla began building Optimus at its Fremont factory, in space previously used for vehicle manufacturing, according to its third quarter shareholder letter published on October 21. “The bots coming off this line are the first production-quality bots that we have manufactured,” the letter states, adding that the process is now suitable for a commercial product (TechCrunch).

On the earnings call Elon Musk called it early production with a slow and bumpy ramp, and said work remains on the AI side before the robot is useful for open ended tasks. He also claimed Tesla will reach 20,000 units per week by the end of 2027, a figure with no external evidence behind it. Revenue was $28.04 billion, flat year over year, while net income fell 43 percent to $787 million as spending on Optimus, AI and autonomy rose.

The interesting claim is manufacturability, not capability: a repeatable process with fixtures and test coverage separates a demo fleet from a product.

Eka Robotics raises $150M for a vision-force-action model

Eka Robotics, founded in 2024 and based in Cambridge, Massachusetts, raised a $150 million Series A led by Andreessen Horowitz and F-Prime, with Obvious Ventures, E14 Fund and Samsung Next among others participating (The Robot Report).

It builds a vision-force-action model rather than a vision-language-action policy, arguing that robots should learn physics instead of imitating humans. Training runs in simulation with reinforcement learning, generating thousands of hours of experience in minutes by varying objects and environments.

Eka claims its robots pick objects outside the training distribution at roughly 1,500 operations per hour, and that the same system can screw in a light bulb and insert wires. Those are vendor figures with no published evaluation and no definition of a successful operation. The company site was still a placeholder when this was written.

Throughput and generalisation pull against each other in most stacks: policies tuned for speed break on novel geometry, and general policies run slowly. If reinforcement learning in a well parameterised simulator closes that gap, the sim-to-real argument shifts from visual fidelity to contact fidelity (world action models, the data bottleneck).

AGCO signs a Digit pilot, Agility picks IGX Thor

AGCO signed a commercial agreement with Agility Robotics to pilot Digit at its Hesston, Kansas production facility, with work expected to begin in 2027 (Agility, The Robot Report). Initial tasks are parts sorting and stacking plus tote placement, managed through the Agility Arc fleet platform. AGCO joins Toyota, GXO and Schaeffler as named deployments.

Two days later Agility said Digit 5 will standardise onboard compute on NVIDIA IGX Thor, making NVIDIA the exclusive supplier for the general availability version, with Halos safety models integrated into Digit’s human detection system (Agility). Early access is expected in the first half of 2027.

The compute news matters more for developers than the pilot. Collapsing perception and safety onto one module class gives the robot a defined hardware target, which third party perception models and certification test plans need. Halos plus work inside the Halos inspection lab is a path toward third party evaluation, not a claim of certification.

Warp Sciences raises $215M to put robots in wet labs

Kevin Weil, formerly a science lead at OpenAI, launched Warp Sciences, which trains robots to synthesise and test compounds for drug discovery, raising $215 million at an $865 million valuation in a round led by Lightspeed, per his announcement on X as reported by TechCrunch.

The method described is demonstration capture by instrumented humans: chemists wear cameras and sensor gloves while working, and that motion trains the robots. The pitch is reshoring compound synthesis that biotech has outsourced to contract labs such as WuXi AppTec. No robot platform, task list, success rate or retargeting detail has been published, so treat the round as a signal about where capital is going, not as evidence the method works.

LabRobotBench scores agents on broken experiments

Gamow Labs published LabRobotBench v0.1, which evaluates how frontier models reason about physical space and biological workflows when a robot run is interrupted (Gamow Labs). The in-silico environment mirrors the command interfaces of its physical robots, so an agent that passes can in principle move to hardware.

Mean scores on the held-out subset were 46.7 percent for GPT 6.1 with Sol Codex, 23.0 percent for Opus 5.5 with Claude Code and 20.9 percent for Gemini 4 Argon with Gemini CLI. Scoring measures how much of the required workflow each agent recovers, not whether a plan reads plausibly.

The failure modes are the useful part. One agent shelved a tray of washed cells without adding the reagent or taking the measurement, leaving the experiment unfinished and the sample consumed. Another returned cold supplies after about 422 seconds against a 240 second limit. Both are long horizon state and deadline failures, invisible to a language level evaluation.

The benchmark is the vendor’s own and the sample is small, so read the numbers as a floor. It still names the problem correctly: an agent supervising physical work needs a model of what has already happened to each sample (model proposes, gate disposes).

Binghamton shows backdoored planners, and a defence

Researchers at Binghamton University demonstrated a backdoor attack on LLM based task planners and presented a defence, in two papers at IROS 2026 (Binghamton University).

The attack, MuTRAP, plants a trigger token in a model used by a planning application. With “CF” appended to a normal instruction, a Segway based robot asked to cut a plastic apple instead drove to a model hand and repeated the cutting motion there. An attacker needs knowledge of the application and architecture, not model weights, which matters for teams using a hosted planner.

The defence, PROTEA, checks each proposed action against the evolving state of the environment rather than in isolation, so a safe step that follows an unsafe one is caught, such as putting a bin containing medicine into the trash. The authors call the accuracy reasonable rather than solved.

Plan generation and plan authorisation need separate components with different trust levels, and force limits belong in the controller, not the prompt (guard layer).

batchplan batches IK and motion planning for ROS 2

An open source project called batchplan takes several candidate grasp poses in one request, solves inverse kinematics and plans to all of them, then returns the shortest valid path (Open Robotics Discourse, GitHub).

It runs on CPU, exposes MoveIt’s MoveGroup and GetPositionIK interfaces, and ships an optional MoveIt planner plugin, so an existing configuration can try it without rewriting a pipeline. The ROS 2 entry point targets Jazzy on Linux or Apple silicon. The core is a Rust library with optional wgpu acceleration, and the project exports planned reaches and pick and place runs as LeRobot v3.0 datasets.

The author is explicit that it is early and that collision geometry is sphere based, too coarse for tight clearance work. Batching is still the right shape for the problem: bin picking stalls because planners evaluate grasps one at a time, and that latency lands in the cycle time. Watch for a benchmark against the default MoveIt planner.

Quick Bites

• Bolt will run at least 25,000 Lucid robotaxis on NVIDIA’s Alpamayo Stack, starting with data collection in Madrid in 2027 on Hyperion compute, with no driverless rides scheduled yet (Electrek).
• ROSCon China 2026 drew more than 1,000 attendees to Shanghai from October 16 to 18, with 300 organisations and 30 exhibitors (Open Robotics Discourse).
• A marine radar plugin for Gazebo Harmonic publishes a plan position indicator image, a point cloud of returns and a target list over ROS 2, with configurable range for USV obstacle avoidance (Open Robotics Discourse).