Force models, backdoored planners, open arms

Eka raised nine figures on a force first manipulation model, Binghamton showed how one word can hijack a robot planner, and open hardware got a mobile manipulator and five hands.

Eka raises $150M on a force first manipulation model

Eka Robotics closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime (TechCrunch). Founders are MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja.

Eka built a vision-force-action model instead of a vision-language-action one: language and force inputs drive the action output through a hand with more force sensors than a conventional gripper. Agrawal argues that vision locates an object roughly, while the last millimeter of a grasp happens through force. Training pairs a physics simulator with reinforcement learning, and the company says it can generate thousands of hours of robot experience in minutes.

The numbers are company claims: about 1,500 objects up to 1 kg picked per hour against roughly 600 for a human, and better than 95 percent reliability. Two outside roboticists told TechCrunch the demo adapted across many objects without retraining.

For developers, force sensing adds cost, calibration and wiring to every fingertip, and contact rich simulation is where simulators are weakest. Watch for an independent evaluation and a task suite outsiders can run. Background: why VLA models matter.

A trigger word turns a robot planner into a weapon

Binghamton University researchers showed a backdoor attack on the large language models that generate robot task plans, with a defense, at IROS 2026 (Binghamton).

The attack, MuTRAP, compromises the planner so a chosen token activates a malicious plan. A mobile robot asked to bring an apple to a cutting board behaves normally. Add the token “CF” and the plan sends it toward a model hand instead. The attacker does not need the weights: a company licensing a third party planner inherits whatever behavior was baked into it.

The defense, PROTEA, reviews each proposed step against the evolving environment. Putting a cardboard box in the trash is harmless, unless the box holds someone’s medicine. That is the gate in front of the model applied to text plans rather than learned policies.

Treat an LLM planner as untrusted. Constrain the action vocabulary, validate each step against preconditions and world state, and log prompt and plan together. The pattern reaches any stack where language meets an actuator.

OpenAMRobot 2.0 publishes the bill of materials

OpenAMRobot posted the specification for version 2.0, an open source dual arm mobile manipulator, with a release candidate due 20 November and a final release on 18 December (Discourse, GitHub).

The hardware specification: two Enactic OpenArm 2.0 arms at 7 degrees of freedom each on a lift column with shoulder height from 1000 to 1350 mm, an NVIDIA Jetson Orin NX, and an STM32H723 running micro-ROS over Ethernet as the base controller. Drive comes from 200 mm ZLTECH hub motors on CANopen, powered by a 25.6 V, 105 Ah LiFePO4 pack. Sensing includes a SLAMTEC RPLIDAR S3, four ultrasonic sensors, two cliff sensors and an IMU. Gross mass stays under 120 kg and software caps speed at 1.5 m/s.

Software is ROS 2 with Nav2, MoveIt 2, ros2_control, a browser operator UI and an embodied AI repository with demonstration recording, an ACT baseline and evaluation. Licensing is MIT, CC BY 4.0 and CERN-OHL-P-2.0 for software, docs and hardware.

Two caveats. The safety layer uses two E-stops that cut motor power through relays while leaving electronics powered, a research grade arrangement rather than a certified one. And no physical robot exists yet: integration and acceptance testing in Berlin runs 23 November to 18 December. The compute matches our Jetson Orin Nano versus Orin NX comparison, and navigation is covered in how Nav2 works.

ROBOTIS opens five hand and gripper designs

ROBOTIS released HX Friends, open source hand and gripper reference designs built on DYNAMIXEL actuators (Discourse, code).

The lineup runs from a 3 degree of freedom single finger with three XM335 actuators and a fingertip tactile sensor, through a 1 degree of freedom gripper on an XM430, to 11 and 12 degree of freedom hands, up to the 20 degree of freedom HX5-D20 five finger hand with five tactile sensors, offered as an option for AI Worker.

The value is the comparison ladder: one actuator family and modular joints from a single finger to 20 degrees of freedom, so a lab can trade parts count against dexterity. Watch whether the ROS 2 packages track upstream releases and whether the tactile sensors ship as standard parts.

A benchmark that scores lab agents on deadlines

Gamow Labs published LabRobotBench v0.1, an in silico benchmark for agents operating wet lab robots (Gamow Labs). It mirrors the company’s physical robot interfaces, so a passing agent has a path to hardware.

The tasks test recovery rather than pick and place. One interrupts a cell assay partway through four trays, one already spoiled, and asks which steps are done without repeating a chemical treatment. Another pits a reagent addition due within 200 seconds against cold supplies that cannot sit out more than 240 seconds. A third requires a measurement within 30 minutes of treatment.

Mean scores were 46.7 percent for GPT 6.1 Sol with Codex, 23.0 percent for Opus 5.5 with Claude Code and 20.9 percent for Gemini 4 Argon with Gemini CLI. The failures instruct more than the ranking: Opus shelved an unfinished assay before adding reagent, Gemini stored cold supplies after about 422 seconds, and GPT reported a completed test whose measurement finished 42.5 seconds late.

The lesson generalizes beyond biology. Benchmarks that score task completion miss the physical constraints that decide whether a result is usable, and a robot reporting success while missing a deadline is worse than one that stops.

Chicago lawsuits test delivery robot biometrics

Two Chicago residents filed class action suits in Cook County court on 9 October against Coco Robotics and Serve Robotics, alleging facial and hand measurements were collected without written consent under the Illinois Biometric Privacy Act (Block Club Chicago).

Both plaintiffs say robot delivery was assigned without their choosing it, through DoorDash and Uber Eats. The suits claim cameras and LiDAR captured face and hand geometry as customers entered codes, and that data was retained for AI training. One points to Serve’s disclosures about licensing datasets; the other claims Coco shared data with OpenAI.

Coco denied collecting biometric data, saying it shares navigational data such as sidewalk hazards with the BlindSquare app. Serve said its systems are not designed to identify individuals or build profiles. Both will defend the suits. BIPA statutory damages run to $1,000 per negligent violation or $5,000 per intentional one. The design lesson is data minimization: a navigation camera does not need to retain faces.

Quick Bites

• Google Cloud and NVIDIA announced general availability next month of A5X instances built on Vera Rubin NVL72, with Google Cloud planning as many as 960,000 Rubin GPUs across sites (NVIDIA). First customer Ineffable Intelligence is using it for large scale reinforcement learning, the workload behind robot policy training.

• A hardware hacker built a stereo vision board around the Rockchip RV1103, a 1.1 GHz single core ARM chip with 64 MB of DDR2, paired with two OV9281 global shutter cameras, an accelerometer and 10/100 Ethernet (Hackaday). Linux and Video4Linux2 support make it a cheap platform for depth experiments.

• Avenda Health added a Surgery Mode to its Unfold AI prostate cancer mapping platform, overlaying 3D risk maps in the surgical view for robotic and non robotic prostatectomies (The Robot Report). It claims 92 percent accuracy for predicting cancer spread against 52 percent for MRI alone, vendor figures to verify.