Robotics Roundup: Living Neurobots, Generalist AI’s GEN-1, and the System Two Bottleneck
Key Definition: Robotics Roundup: Living Neurobots, Generalist AI’s GEN-1, and the System Two Bottleneck is [add clear definition here].
The robotics world is moving fast — and in two directions at once. On one side, embodied AI models are crossing new performance thresholds in physical tasks. On the other, researchers are building robots out of living cells that grow their own nervous systems. Meanwhile, one of the field’s most respected voices is reminding everyone that the “brain” problem still isn’t solved.
Here’s your weekly roundup.
Generalist AI Launches GEN-1: 99% Success Rate on Physical Tasks
Generalist AI has introduced GEN-1, an embodied foundation model designed to perceive, reason, and act in the physical world — and the company claims it crosses a new performance threshold: mastery of simple physical tasks.
The numbers are striking. GEN-1 achieves 99% success rates on certain tasks, compared to ~64% for its predecessor GEN-0, while completing tasks up to three times faster. Perhaps most notably, it requires only roughly one hour of robot-specific data to adapt to new tasks — a dramatic reduction from the expensive teleoperation datasets traditionally required.
How It Works
GEN-1 is trained on large-scale pretraining of human activity data collected through wearable devices, rather than relying solely on teleoperation. This mirrors the LLM playbook: use the cheapest, most abundant data source available.
The company defines “mastery” as a combination of three factors:
- Reliability — consistent performance over hundreds or thousands of repetitions
- Speed — completing tasks efficiently, not just correctly
- Improvisational intelligence — the ability to adapt when unexpected situations arise
Demonstrations show robots folding boxes, packing items, and assembling components over extended periods with minimal errors — sometimes completing thousands of repetitions.
The Scaling Law Question
GEN-0 demonstrated that “scaling laws exist in robotics,” drawing a parallel with LLMs. GEN-1 extends this by increasing both data and compute while introducing new training and inference techniques. If the scaling law holds in the physical world, we may be at a GPT-2-to-GPT-4 inflection point for embodied AI.
The company acknowledges limitations — not all tasks reach production-level performance yet — and early access is limited to selected partners.
Gill Pratt on Humanoid Robots: The Body Hasn’t Changed, the Brain Has
Gill Pratt — architect of the DARPA Robotics Challenge and current CEO of the Toyota Research Institute — spoke with IEEE Spectrum about the state of humanoid robotics, ten years after the DRC finals.
His core argument: what has changed isn’t the body, it’s the brain.
System One vs. System Two
Pratt frames the current state of robot AI through Kahneman’s dual-process theory:
- System One (fast, reflexive pattern matching): Diffusion policies and Large Behavior Models represent extraordinary progress here. TRI’s own diffusion policy work kicked off an entire field — “every robotics demonstration we’ve seen is using some form of diffusion policy.”
- System Two (slow, deliberative reasoning): Still missing. “Patching System One to make it System Two is like squeezing a water balloon — you fix one thing and something else breaks.”
His proposed near-term solution mirrors autonomous driving: robots operate autonomously most of the time, but call home for human help when they encounter edge cases — essentially using humans as System Two providers.
On the Humanoid Hype
Pratt is measured but pointed:
- Legs aren’t always optimal. “It’s very weird to see so much focus on legged robots in factories, which are flat environments perfectly suited for wheels.”
- But the form factor has advantages: the world is built for human bodies (physical affordances), and imitation learning works better when the robot shares our morphology.
- The real opportunity is elder care. TRI is exploring “care-receiving robots” — robots that humans teach by demonstration, giving the teacher a sense of purpose. “We have evolved to be creatures that love giving and helping.”
His bottom line: “We’ve had incredible breakthroughs with System One, but it doesn’t mean the robots are going to be doing all that much, unless somebody makes a System Two breakthrough also.”
Living Neurobots: When Robots Grow Their Own Nervous Systems
Researchers at Tufts University, led by Michael Levin, have created what they call neurobots — biological machines built from frog cells that spontaneously develop functional neural circuits, as reported in Advanced Science.
From Xenobots to Neurobots
| Generation | Year | Control Mechanism | | Xenobots | 2020 | Physical/anatomical (cilia-driven swimming, self-repair, self-replication) | | Neurobots | 2026 | Internal nervous system (electrochemical signaling between neurons) |
The key advance: neurons mature from partially differentiated stem cells alongside structural tissues, forming branching connections throughout the organism. This allows neurobots to relay signals from cell to cell — enabling a layer of internal control that earlier xenobots lacked entirely.
Observable Differences
Neurobots idle less, explore more, and trace looping and spiraling paths (rather than repeating simple trajectories). They also respond differently to neuroactive drugs — the first time electrical signaling has been directly linked to observable movement in a biological machine.
What’s Next
The team plans to add human neural cells to “anthrobots” (built from human lung cells), extending the neurobot framework into a fully human context. Commercial startup Fauna Systems is initially targeting environmental sensing — deploying xenobots in aquaculture, wastewater monitoring, and pollutant detection.
As Levin puts it: “Where does form and function come from in the first place? When it’s not evolved and it’s not engineered, where do these patterns come from?”
CorTec BCI Receives FDA Breakthrough Designation for Stroke Rehabilitation
CorTec’s Brain Interchange system has received FDA “Breakthrough Device Designation” for use in stroke motor rehabilitation — the first BCI globally to receive this recognition specifically for stroke.
The Technology
Unlike Neuralink, Synchron, or Blackrock Neurotech — which focus on enabling communication through external devices — CorTec’s system is designed to restore motor function through:
- Fully implantable, wireless, bidirectional closed-loop architecture
- Simultaneous neural signal recording + adaptive electrical stimulation
- Signal stability demonstrated over 500+ days (published in Nature Scientific Data)
Clinical Status
An FDA-approved IDE study is underway at the University of Washington in Seattle — the first human clinical trial of a fully implantable wireless BCI for stroke rehabilitation.
The company is also exploring applications in epilepsy, paralysis, and depression.
ABB × Jacobi: AI-Powered Mixed-Case Palletizing at Scale
ABB Robotics and Jacobi Robotics have partnered to integrate Jacobi’s OmniPalletizer AI software into ABB’s global robotics ecosystem — bringing AI-powered mixed-case palletizing to the integrator channel at scale.
Mixed-case palletizing — building stable, store-ready pallets from unsequenced case flows — costs the industry over $15 billion per year in direct labor in the US alone (and $50+ billion including indirect costs). It has remained one of the last major warehouse workflows resistant to automation.
Key advantages of the collaboration:
- Brownfield-ready: Drops into existing conveyor lanes without sequencing infrastructure
- Digital twin validation: Every deployment is pre-validated against the customer’s actual order history
- Fleet-wide learning: Stacking performance improves continuously across all deployed sites
Live demo at MODEX 2026 in Atlanta, April 13–16.
Also Worth Noting
Logic Octopus: Ceiling-Mounted Multi-Arm Warehouse Robot
Logic has unveiled Octopus, an overhead picking robot with configurable multi-arm setup (suction cups, clamps, grippers, specialized end effectors — all live simultaneously). Paired with autonomous Logic Pallets that navigate goods to the right arm, it implements a true goods-to-robot paradigm without consuming floor space.
PickNik MoveIt Pro 9: Scan-and-Plan for Unstructured Environments
PickNik released MoveIt Pro 9, strengthening scan-and-plan workflows that let robots perceive their surroundings in real time and generate motion paths dynamically. Already deployed in vehicle washing (Autowash), restroom cleaning (Hivebotics), and food plant sanitation (CleanBotix).
Tennant X16 Sweep: Autonomous Industrial Cleaning
Tennant launched the X16 Sweep, its first autonomous robotic sweeper powered by Brain Corp’s BrainOS. Features SelfPath AI for dynamic path generation, DustShield for dusty environments, and optional auto-docking for 24/7 multi-shift operation.
Radiation-Hardened Wi-Fi for Nuclear Decommissioning
Researchers at Institute of Science Tokyo have built a 2.4 GHz Wi-Fi receiver that survives 500 kGy of radiation — over 1,000× what space-rated electronics endure, and 3,000× more than a KUKA robot arm could handle before failing. The goal: wireless control for robots decommissioning nuclear reactors, eliminating the tangled LAN cables that plagued Fukushima cleanup efforts.
RAI Institute: 10,000 People Drove Spot
RAI Institute ran a free public robot experience at CambridgeSide mall where approximately 10,000 guests visited and over 1,000 drove a Boston Dynamics Spot. Key finding: comfort scores increased significantly across all contexts after just a few minutes of hands-on control — the largest gains were in outdoor/disaster scenarios. 74% reported “excitement” post-interaction; only 12% reported nervousness.
The Takeaway
Three themes emerged this week:
-
The embodied AI scaling law is real, but uneven. GEN-1’s 99% success rate is impressive, but Generalist AI itself admits not all tasks reach production level yet. We’re on the curve, but the curve has a long tail.
-
The System Two gap is the defining constraint. Gill Pratt’s framing is the clearest articulation yet: diffusion policies solved System One for robotics, but without imagination, planning, and true reasoning, robots will remain reactive. The near-term answer — humans as remote System Two — is practical but feels like a bridge, not a destination.
-
Biology is entering the stack. Neurobots aren’t a curiosity — they represent a fundamentally different approach to building autonomous systems. When your robot grows its own nervous system, the line between engineered and alive becomes a question, not a fact.
Frequently Asked Questions
What is Robotics Roundup: Living Neurobots, Generalist AI’s GEN-1, and the System Two Bottleneck?
[Provide a direct answer in 40-60 words that can stand alone as a complete response.]
How does Robotics Roundup: Living Neurobots, Generalist AI’s GEN-1, and the System Two Bottleneck work?
[Explain the mechanism or process clearly, using numbered steps if applicable.]
What are the main risks or challenges?
[Provide a balanced assessment of limitations and obstacles.]
Sources: IEEE Spectrum, Robotics and Automation News, Advanced Science, Generalist AI, ISSCC 2026
GEO optimized: 2026-05-23