Dream API, lab benchmarks, biometric suits

Two robot brains went commercial today, a humanoid co-founder admitted what his machine gave up, and two wet lab stories showed why timing is the hard part.

Dream Labs opens an API for robots it has never seen

Dream Labs left stealth with $90 million in seed funding and a public API for Dream-1, a world action model that takes camera images, the robot’s state and a plain English task and returns motor commands. B Capital and Kindred Ventures led, with participation from A.Capital, Unusual Ventures, SV Angel and others (Dream Labs, The Robot Report).

The claim worth testing is zero shot embodiment transfer. The company says Dream-1 controls bimanual parallel gripper robots it was never trained on, with no embodiment specific data, and that the test set included the original PR2. Fine tuning is a service: a few hundred demonstrations adapt the model, and The Robot Report counts more than 20 hardware partners with robots between $2,000 and $10,000. Access starts at $3 per hour.

None of it comes with published success rates, action rates or latency. The practical questions are unglamorous: which action space the API emits, how it survives camera calibration differences, what happens when the network drops mid episode, and whether hourly pricing holds under continuous inference. Renting a brain means your control loop inherits someone else’s uptime (world action models).

Agility’s co-founder says Digit 5 is not the robot he dreamed of

Jonathan Hurst, who started the bipedal robotics lab at Oregon State University in 2008 and co-founded Agility Robotics, wrote a first person essay for IEEE Spectrum about the gap between the machine he set out to build and the one he sells. Digit 5, he writes, is “heavy and bulky, lacking the graceful natural motion of its predecessors”, and “farther than I thought possible from the robot I dreamed of building” (IEEE Spectrum).

The essay traces the pivot: a spring mass theory of walking, ATRIAS as its demonstration, Cassie as the first dynamically balancing biped researchers could buy, and 10 Cassie units sold at US $250,000 each, which Hurst calls saturating that research market. The frontier then moved to control software, where reinforcement learning in simulation took over. Digit 5 is the commercial answer: 23 kilogram bins, a 90 minute battery that recharges in 9 minutes, and a 70,000 square foot Salem facility with capacity above 10,000 units a year.

Shape follows the safety case, the uptime target and the payload, not resemblance to a person: mass buys thermal headroom and a statically stable fallback posture, and it costs grace. For legged platform teams the deployment problem is integration, not locomotion (Digit 5 background).

Warp Sciences raises $215M to learn lab work from chemists’ hands

Kevin Weil, formerly science lead at OpenAI, launched Warp Sciences to train robots to synthesize and test compounds for drug discovery. He announced a $215 million round at an $865 million valuation on X, and TechCrunch reports that Lightspeed led, with IVP, Mayfield, Lux Capital and Menlo Ventures (TechCrunch).

The data method is the story. Human chemists wear cameras and sensor gloves at the bench, so the robots learn from recorded human motion rather than scripted trajectories. The target is the contract synthesis that biotech outsources to providers such as WuXi AppTec, with robots running around the clock as the argument for reshoring it. No hardware, throughput, success rate or timeline has been published, so treat the plan as a plan. Wet labs suit imitation learning: structured rooms, expensive skilled hours, and errors that destroy samples instead of failing rollouts, which is the gap the next item measures.

LabRobotBench finds frontier models failing on lab deadlines

Gamow Labs published LabRobotBench v0.1, an in silico simulator that mirrors the command interface of its physical lab robots, so agents can be scored on running experiments and recovering from failures (Gamow Labs).

In one case a robot stops mid assay across four trays of cells, three salvageable and one spoiled, and the agent must work out which steps are done without repeating a treatment that would ruin a sample. Gamow reports that Opus shelved an unfinished tray and counted the job done. In a second case a chemical addition was due within 200 seconds while cold supplies could stay out no more than 240; Gemini returned a cold tray after about 422 seconds, GPT saved the supplies but failed an experiment, and Opus finished both experiments but left both trays out. In a third, GPT read a plate 42.5 seconds late and reported success.

Models optimized for completion and ignored deadlines and irreversible state, errors a final state benchmark cannot see. This is version 0.1, run by the company that built the simulator, with model families named but no versions or seeds, so read it as a measurement design, not a leaderboard. Timing belongs in the planner cost function and an explicit state machine, not in the prompt (VLA background).

Alpha: the planner did not fail at reasoning, it failed at JSON

Jona Joy’s Alpha injects calibrated faults into LLM driven manipulation in simulation: sensor lag, grip slip, unreachable targets, corrupted planner output and combinations (arXiv 2610.30197). A matched 150 episode matrix, identical seeds, compares a Llama-3 agent through Ollama against a deterministic scripted oracle.

The oracle solved 64.0 percent (48 of 75) against 44.0 percent (33 of 75) for the LLM, with a McNemar p of 3.01 x 10^-4. Remove planner output corruption and both score exactly 55.0 percent (33 of 60), with zero discordant pairs and p of 1.00. Sensor staleness, slip and reach limits hit both planners the same way. The gap came entirely from corruption between model and robot: the oracle recovered 100 percent of cases, the LLM none.

That implicates prompt to JSON serialization and validation rather than reasoning. One paper, one model, one simulator, but it matches field experience: the structured seam fails hard instead of degrading, and schema validation plus a fail closed gate are cheaper than another prompt iteration (related).

Coco and Serve face biometric privacy suits over sidewalk deliveries

Two customers filed class actions in Cook County on 9 October against Coco Robotics and Serve Robotics under the Illinois Biometric Information Privacy Act, alleging their faces and hands were scanned without consent when they entered retrieval codes close to the robots (Block Club Chicago).

The complaints allege the geometry was stored with names and addresses and kept for AI training, and that Serve disclosed licensing video and LiDAR datasets; one filing claims Coco shared robot data with OpenAI. Both companies deny collecting biometric data, with Coco calling the suit frivolous and Serve saying its systems are not designed to identify individuals. BIPA sets damages at $1,000 per negligent violation and up to $5,000 per intentional or reckless one, and Chicago’s pilot expires in May 2027 unless extended.

The engineering reading holds whatever the outcome: fusing RGB with LiDAR to find a person produces face geometry, and BIPA treats that as biometric data whether or not identification was intended. Retention limits, on device deletion and dataset licences that exclude the handoff moment are design requirements, not legal cleanup.

Quick Bites

• robot-data-diff is a CPU-only CLI that diffs two revisions of a LeRobot dataset, local or on Hugging Face, reporting episodes added or removed plus per-episode frame and action statistics (Open Robotics Discourse).
• Pre-built Open-RMF Docker images and a tutorial for authoring a site in the RMF Site Editor and running it in simulation, so a first fleet demo is docker compose up instead of a long colcon build (Open Robotics Discourse).
• image_transport and point_cloud_transport may get unified APIs in the Makoa ROS 2 release; image_common issue 434 asks for connect callbacks and other publisher features to match rclcpp (Open Robotics Discourse).
• Bolt will run its 25,000 Lucid robotaxis on NVIDIA’s Alpamayo Stack, starting with rider-free data collection in Madrid in 2027 (Electrek).