Frontier MechanicsField report / 028
Frontier Mechanics Week 28: Helix controls two robots, NEO enters the home, Karman mines, and Veo prices video
Figure fused language and action, 1X revealed a domestic humanoid, an asteroid startup raised $20 million, and generated video acquired a per-second meter.
Figure introduced Helix, a vision-language-action model controlling a humanoid’s upper body and, in one demonstration, coordinating two robots. 1X revealed NEO Gamma for domestic testing. Karman+ raised $20 million for an autonomous asteroid mission, and Google priced Veo 2 video generation at fifty cents per second.
The week converted impressive capabilities into units a business can manage: control frequency, household tasks, mission stages, and generated seconds.
Figure’s Helix coordinated two robots, 1X aimed NEO at the home, Karman+ financed asteroid reconnaissance, and Google priced Veo 2 by generated second. This was a week about scarce experience. Helix needed cross-robot demonstrations. NEO needed data from messy houses. Karman needed to observe an asteroid before designing a mine. Veo needed enough compute and training to produce a usable clip. The business question was who pays for learning before the system becomes reliable enough to sell.
Robotics models
Helix splits slow intent from fast motion
Figure’s Helix combines a larger vision-language layer that interprets scenes and instructions with a faster visuomotor policy producing continuous upper-body actions. The system controlled wrists, torso, head, and fingers, handled unfamiliar household objects, and coordinated two robots on a shared task using one set of weights.
The layered rates matter. Semantic planning can run slowly; balance and contact cannot wait for a paragraph. Helix’s promise is not that language directly twitches every motor, but that learned representations pass goals into a policy fast enough for manipulation. Demonstrations remain curated and teleoperation history matters. Still, the architecture looks like a plausible division of cognitive labor.
Helix used a vision-language-action architecture with a slow semantic system and a fast visuomotor policy. The higher layer interpreted language and scene, choosing what objects and goals mattered. The lower layer produced high-rate joint actions. This split matched the clocks of intelligence: intent can update a few times per second; balance and manipulation need much faster feedback.
Figure demonstrated two humanoids collaborating to place groceries, responding to natural-language prompts and unfamiliar objects. Coordination required each robot to perceive shared space and avoid conflicts. Central planning can assign roles; decentralized policies react locally. Communication delay and uncertainty make explicit task state important.
A shared policy across two robots increases data efficiency and common-mode risk. If both misclassify the same object or choose symmetric paths, they can fail together. Diversity in sensing, role, or policy may improve resilience. Multi-robot benchmarks should include recovery when one agent stops or provides bad state.
The slow–fast architecture also supports safety layering. The semantic model can request an action; the fast controller enforces kinematics and contact. Language should not directly emit raw torque. The interface between layers needs an action vocabulary expressive enough for generality and constrained enough for verification.
Helix’s significance was deployment on the robot without a scripted object list. The remaining proof was rate and robustness: long sequences, clutter, force, falls, battery, and human proximity. Coordination makes a demo richer and fleet operations harder. Two robots do not double throughput if they spend time negotiating the same narrow workspace.
Helix's split between slow semantic planning and fast visuomotor control mirrors robotics' timescale hierarchy. A high-level model decides which object and goal matter at perhaps a few hertz; a lower policy converts current images and joint state into continuous actions much faster. This reduces the cost of running a large model at servo rate and lets reflexive correction continue while language-level reasoning remains stable.
The boundary can fail when the slow plan becomes stale. An object moves, a grasp slips, or a person enters the workspace while the semantic controller still believes the old scene. The fast policy needs authority to pause and request replanning, plus a representation that communicates failure rather than force completion. Latency budgets should include perception, inference, network transfer, and actuator response under peak load.
Generalization claims should separate vocabulary from dynamics. Recognizing a novel object and selecting an appropriate grasp are different achievements; transferring between robot bodies adds calibration, reachability, and force limits. Evaluation needs unseen objects, clutter, changed lighting, and controlled perturbations, with success and recovery measured over full tasks. A two-speed model is compelling when it fails legibly across these boundaries.
Home robotics
NEO volunteers for the messy house
1X revealed NEO Gamma, a soft-covered humanoid intended for home testing. The design emphasizes quiet actuators, human-scale motion, and physical compliance. Unlike a warehouse, a home contains pets, narrow passages, glassware, loose fabric, and people who did not complete a safety induction.
Domestic deployment can generate the varied data general robots need, but it makes users part of the research environment. Teleoperation may help the machine complete tasks and collect demonstrations, raising privacy and labor questions. A home robot must be safe when confused, useful when supervised, and transparent about when a distant human is looking through its eyes. That is a taller order than the robot.
1X’s NEO targeted household tasks with a soft-bodied humanoid form. Homes reward human scale and punish industrial assumptions. Floors change, objects move, children and pets enter, and no safety engineer marks a cage. Soft covers and lower force can reduce injury while also reducing payload and precision. The machine has to be safe before it is useful enough.
Housework includes deformable objects, drawers, dishes, liquids, and social preferences. Success is not only placing an item but placing it where this household expects. Learning requires demonstrations and correction from people who are not roboticists. Teleoperation can bootstrap tasks and raises privacy questions because remote operators may see inside homes.
A home robot needs an honest autonomy label. If a person assists remotely, users should know when and what data leave the house. Connectivity loss should produce a safe pause. Tasks involving heat, knives, medication, doors, or children need strict exclusions until evidence supports them. A general body should not imply general authority.
Economics differ from factories. Household willingness to pay is limited, utilization may be low, and support visits expensive. Leasing can spread hardware cost and bundle teleoperation. The provider then carries fleet capital and labor. A useful early market may be structured homes, elder assistance under supervision, or repetitive chores with measurable value.
NEO’s challenge was to accumulate real-world data without turning early customers into unconsented lab staff. Local learning, clear capture controls, and benefit for teaching matter. The home is a rich dataset because it is intimate. That is the reason to protect it, not a reason to claim implied permission.
A messy home punishes scripted demonstrations because the distribution changes between runs. Chairs move, laundry deforms, transparent objects disappear from depth sensors, and people create partial goals. NEO needs a belief state that preserves uncertainty about object identity and location rather than converting every frame into a confident scene graph. Active perception—moving the camera, touching gently, or asking—should be part of the policy.
Remote assistance can accelerate deployment but must be measured honestly. A human who labels an object, selects a grasp, or directly drives the arm is supplying labor and safety supervision. Report assisted-task fraction, intervention duration, latency, and whether the correction becomes reusable learning. The economics can still work, but only if assistance declines or is cheaper than the household labor displaced.
The household safety case is social as well as mechanical. Children and pets violate expected trajectories; private rooms and personal objects carry permissions; a harmless rearrangement can still be unwelcome. Local emergency stop, force and speed limits, secure video handling, and clear activity logs are basic. A robot invited into a home must be easier to audit and correct than the cloud service behind it.
Deep space
Karman+ starts with reconnaissance
Karman+ raised $20 million to develop High Frontier, an autonomous spacecraft intended to reach a near-Earth asteroid, characterize it, and eventually extract material. Asteroid mining’s ultimate revenue lies beyond several technical businesses: navigation, power, autonomy, sampling, processing, and transporting something useful.
The company sensibly begins by buying information. A reconnaissance mission can validate the spacecraft and reduce uncertainty about target composition. Investors are financing staged options rather than immediate commodity flow. The asteroid is not the first product. The first product is evidence that the next mission deserves to exist.
Karman+ planned small spacecraft to rendezvous with near-Earth asteroids, characterize resources, and eventually support extraction. Reconnaissance comes first because asteroid size, composition, spin, surface cohesion, and orbit determine every later design. A mine built from telescope estimates is a collection of unpriced unknowns.
Small bodies have weak gravity. A spacecraft cannot rely on weight for traction, and material disturbed by drilling may escape. Rotation complicates approach and anchoring. Some asteroids are rubble piles rather than coherent rock. Imaging, spectroscopy, and radio science can narrow these variables before contact hardware exists.
Autonomy is necessary because round-trip light delay prevents joystick control. The vehicle must navigate against a dark irregular target, manage approach, and enter safe mode without immediate help. Optical navigation compares target position and stars; relative motion near a low-gravity body is sensitive to small forces.
The $20 million raise financed a sequence of missions rather than a mine. Each reconnaissance vehicle can build a catalog and flight heritage. Revenue before extraction might come from data, government science, or strategic resource surveys. Without intermediate customers, investors finance several missions before the thesis earns cash.
Resource economics depend on use location. Returning bulk metal to Earth competes with terrestrial mining and reentry. Water used as propellant or life support in space avoids launch cost and can be more valuable. The market requires vehicles and habitats that consume it. Prospecting is therefore scientific work and demand creation for an economy not yet synchronized.
Asteroid mining begins with characterization because resource value is highly heterogeneous. Spectra suggest composition at the surface, while radar, thermal inertia, rotation, and gravity constrain structure and accessibility. A metal-rich classification does not reveal whether useful material is concentrated, buried, fractured, or economically separable. Reconnaissance should update a distribution over extraction outcomes, not produce a single promotional tonnage.
Karman+'s vehicle must trade scientific coverage against propulsion. Every flyby geometry or proximity maneuver spends time and delta-v while navigation around a small irregular body is uncertain. Weak gravity makes anchoring and excavation reaction forces difficult; ejecta can escape or contaminate equipment. Early missions should measure regolith mechanics and surface interaction because those properties determine whether any processing architecture closes.
The market endpoint also matters. Returning platinum to Earth competes with terrestrial supply and can depress its own price; using water or metal in space requires customers whose missions are designed to buy it. Discounted cash flow across long development and travel times is brutal. Reconnaissance creates value if the data supports a tradable resource claim or mission decision long before a refinery is promised.
Generative media
Veo charges by the finished second
Google priced Veo 2 generation through its API at fifty cents per second of output, or $30 for a minute before retries and editing. A price per second makes video generation resemble cloud rendering rather than a subscription feature. It also exposes the cost of exploration: creators rarely keep the first sample.
The economic comparison is not simply with filming a minute of footage. Generated video competes with stock, previsualization, visual effects, and scenes that would not otherwise be produced. Lower prices will increase attempts, so total compute can rise as unit cost falls. The meter changes behavior: every prompt becomes a tiny production decision.
Google priced Veo 2 at fifty cents per generated second through its API, making a minute of output cost thirty dollars before retries and editing. Video generation is expensive because the model must maintain spatial and temporal coherence across many frames. The price converted an impressive demo into a production budget.
A creator rarely uses the first generation. Prompt iterations, alternate takes, extensions, and failed clips multiply cost. The relevant unit is accepted second, not generated second. If one in five clips is usable, effective model cost rises fivefold. Editing can salvage partial output and improve the ratio.
Video models use compressed latent representations to avoid operating directly on every pixel, then decode frames. Temporal attention or related mechanisms preserve motion and identity. Longer clips increase memory and the chance of drift: objects change shape, physics breaks, and cuts appear. Cost scales with resolution, duration, sampling, and model steps.
API pricing creates product segmentation. Advertising, previsualization, and effects may tolerate tens of dollars per minute because alternatives involve crews and rendering. Casual social video may not. Batch or lower-quality tiers can broaden use. Customers pay for controllability, rights, consistency, and speed as much as visual fidelity.
The price also implied a large infrastructure market and an incentive for efficiency. Distillation, caching, faster samplers, and specialized hardware would lower cost. As generation gets cheaper, output volume rises and provenance becomes more important. The accepted second should carry metadata about model, edit history, and source policy.
Charging by finished video second hides a large variance in compute. A short prompt with simple camera motion may converge quickly; a precise scene with characters, text, and edits may require several generations. The provider can average this variance across users, while creators experience it as iteration cost. A useful meter would show generation attempts, resolution, duration, and compute tier alongside the exported seconds.
Video generation also moves labor from production into specification and selection. The user must describe continuity, camera, motion, lighting, and exclusions, then identify failure across frames. Temporal artifacts are expensive because one bad hand or object identity can spoil the clip. Editing controls, reference locking, and localized regeneration may reduce total cost more than another gain in raw visual fidelity.
The rights layer becomes part of unit economics. Training provenance, likeness consent, brand use, and output indemnity influence whether a generated second can enter an advertisement or film. Cheap pixels are not equivalent to usable footage. Veo's durable product is a workflow where a customer can predict both rendering cost and legal/creative revision cost before committing the rest of a production around the result.
Helix, NEO, Karman+, and Veo all make generative systems operational by giving them a rate or stage. The poetry of general intelligence becomes a control loop at a particular frequency, a supervised home trial, a reconnaissance flight, or thirty dollars per finished minute. Progress becomes investable when somebody finds the unit.
Learning curves need a payer. Figure can fund data through industrial ambition, 1X through early households and investors, Karman through mission capital, and Google through per-second API revenue. The choice shapes which data arrive.
A factory rewards throughput, a home rewards fit and privacy, an asteroid mission rewards uncertainty reduction, and a video API rewards accepted creative output. Treating them all as scaling problems erases the objective that should govern the dataset.
Early users provide more than revenue. They supply edge cases, preferences, and evidence. Companies owe them explicit compensation or product value, plus boundaries around how that experience trains future systems. ‘Beta’ should not be a waiver of responsibility.
The marginal example is valuable only if the learning loop can absorb it. Logs need outcome labels; failures need taxonomy; model updates need tests. Data accumulation without curation is storage, not progress.
Scarce experience is why frontier systems improve unevenly. The fastest route is not always gathering more. It is placing the system where every failure answers a question the team is prepared to ask.