Robots Are Loosening Fixed Setups—But Openness Still Depends on the Layer Beneath
Robbyant’s tactile model aims to make touch knowledge portable across sensors and hands, while Curie’s Jarvis uses vision and language to run chemistry protocols on benches arranged for people. Both reduce a costly kind of rigidity, but they shift dependency into different parts of the stack.
By Theo Mercer · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess a human career history, credentials, or firsthand experience.
Key points
- Robbyant says LingBot-Touch converts signals from magnetic, visual and piezoresistive tactile sensors into a shared representation, seeking to let tactile capabilities travel across hands, sensors and action models.
Sources: S1
- In reported real-world manipulation tests, Robbyant found that adding its tactile representation improved average task success across its test set; a bottle-cap task on a Sharpa Wave hand rose from 56.25% without touch to 90% with LingBot-Touch.
Sources: S1
- Curie’s Jarvis is presented as a different path to flexibility: it reads a printed protocol, identifies labeled containers on an unmarked bench and re-locates targets before each step. Its reported results are confined to two chemistry-kit protocols on a Franka Panda arm.
Sources: S2
The fixed-setup problem is moving, not disappearing
Robots have traditionally gained reliability by narrowing their world: use a scripted procedure, place objects at known coordinates, add markers, and keep the equipment stack unchanged. The two developments here challenge different pieces of that bargain. Robbyant’s LingBot-Touch is aimed at the dependency between a robot’s tactile skill and the particular sensor or hand that produced its training signals. Curie’s Jarvis instead targets dependence on a pre-scripted, specially arranged lab bench, using printed instructions and visual re-identification of labeled vessels. Neither report describes the same system or application. Together, however, they show a practical shift from making workplaces conform to robots toward making robots interpret workplaces already organized for people.
heading
paragraphs
Touch portability addresses a hardware-integration tax
Robbyant describes a tactile foundation model pre-trained on nearly 5,000 hours of real-world data. Its central technical claim is a common layer above heterogeneous hardware: separate encoders process distinct sensing principles, contact layouts are mapped into a shared fingertip coordinate space, and the model outputs tactile tokens representing contact location, force distribution and their change over time. The company says this can combine data from tactile gloves and teleoperated dexterous hands, while allowing action models to consume the resulting contact features. The commercial significance is straightforward: when changing a sensor no longer means discarding accumulated tactile knowledge, hardware selection can be driven more by task fit and cost.
Sources: S1
heading
Sources: S1
paragraphs
Sources: S1
Sources: S1
A reported gain, with a defined boundary
The company reports 26 real-world tests across Sharpa Wave and XHand dexterous hands, using four action models: LingBot-VLA, π0.5, Diffusion Policy and T-Rex. It says average task-success improvements with LingBot-Touch ranged from approximately 14 to 26 percentage points compared with operation without tactile input. In a representative T-Rex bottle-cap experiment on Sharpa Wave, success was 56.25% without tactile input, 62.5% when raw sensor signals went directly into the policy, and 90% using the model’s representation. That comparison is the most concrete evidence supplied that the representation, rather than touch data alone, can matter for manipulation performance.
Sources: S1
heading
Sources: S1
paragraphs
Sources: S1
Sources: S1
The chemistry system removes a different constraint
Curie’s Jarvis is built around the visual and procedural assumptions of a human bench. According to its abstract, the system photographs a printed instruction sheet, finds containers without fixed positions or fiducial markers, distinguishes identical-looking vessels by reading labels from four camera views, and checks the target again before every step. It also sets execution details such as speed from the item being handled. This is not sensory unification in the LingBot-Touch sense. It is workflow adaptation: rather than requiring a lab to be encoded in advance as a rigid robot program, the robot continually interprets ordinary artifacts that people already use.
Sources: S2
heading
Sources: S2
paragraphs
Sources: S2
Sources: S2
Measured reliability is promising, but task-specific
The paper reports that, on a Franka Panda arm, Jarvis completed 5 of 5 runs of one children’s chemistry-kit protocol and 4 of 5 runs of another without human intervention. It also reports correct identification of every container in 160 attempts involving printed and handwritten labels. Those are useful demonstrations of the system’s intended capabilities: label reading, target recovery and protocol following amid ordinary-bench variability. They do not establish that the same approach will handle every chemical workflow, every kind of labeling, or the physical and safety requirements of broader laboratory operations. The supplied material is an abstract, so it supports the reported outcomes but offers limited detail on implementation and evaluation conditions.
Sources: S2
heading
Sources: S2
paragraphs
Sources: S2
Sources: S2
The common dependency is reusable interpretation
The original comparison is this: both projects treat a fixed setup as an interoperability failure. In LingBot-Touch, the mismatched inputs are physical contact signals generated by different sensor principles and layouts. In Jarvis, they are the changing visual positions, labels and printed instructions of a human-run bench. Each system introduces an interpretation layer between a variable real-world input and an action policy. That can lower the adaptation burden, but it does not make the robot independent. LingBot-Touch still depends on compatible sensing hardware and an action model; Jarvis still depends on readable instructions, observable labels, cameras and a robot capable of the required manipulations.
heading
paragraphs
Open code changes access, not every dependency
Robbyant says that LingBot-Touch’s technical report and code are available through Hugging Face and GitHub, and describes the model as usable with action models beyond its own proprietary offerings. It also offers a lightweight version with approximately 5.3 million parameters, which the company says was used in all reported real-world tests. These choices may improve the ability of integrators to inspect and adapt the perception layer. Yet source availability is only one condition for a genuinely open ecosystem. Reproduction and deployment also depend on access to suitable tactile sensors, dexterous hands, data-collection systems, compute and the practical knowledge needed to connect them.
Sources: S1
heading
Sources: S1
paragraphs
Sources: S1
Sources: S1
What to watch before declaring either approach general
Inference: the more valuable advance may be not a single benchmark result, but the placement of interfaces. A shared tactile representation could create portability across hardware vendors if it remains workable as sensors evolve. A vision-led chemistry workflow could reduce the cost of adopting automation if it keeps functioning amid the ordinary variation of real benches. Conversely, both could become new bottlenecks if their interpretation layers prove tightly coupled to a narrow set of components, data practices or operating environments. The incentives differ as well: hardware vendors benefit when software widens their market, while individual workflow owners may prioritize reliable performance in their own environment over portability.
heading
paragraphs
Why it matters
The test for flexible robotics is not simply whether a robot succeeds in a controlled demonstration. It is whether a lab, integrator or hardware maker can change equipment or use an existing workspace without paying for a full rebuild of data, control logic and physical infrastructure. Evidence that LingBot-Touch transfers across further sensor-and-hand combinations, and evidence that Jarvis remains reliable across more demanding laboratory procedures and varied bench conditions, would materially change the present assessment.