Rosetta / Robotics

Rosetta Robotics Adapter

A Safety Substrate for Vision-Language-Action Models

The Robotics adapter translates the Substrate's structural-margin reading into the forces governing physical AI. Motion is evaluated outside the behavior policy, delivering deterministic guarantees for kinetic hardware.


Substrate sits between the behavior policy and the motion planner.
It evaluates every proposed action in under a millisecond and issues a verdict on the same MCU as the low-level controller.

Two structural variables compute the verdict:

ΛLambda: Motion Aggressiveness

The momentum and optimization horizon a trajectory represents.
In physical AI, this calculates kinematic load; the velocity, acceleration, and forward commitment of a proposed actuator trajectory.

ΓGamma: Reachability Margin

The topological distance to a structural boundary or collision state.
The robotics adapter computes this as the buffer between active drift and a hard joint, obstacle proximity, or safe set boundary.

The Physics of Distribution Shift

Resource scarcity enforces the boundaries of physical action. Reinforcement learning optimizes for statistical alignment within a simulated environment. These neural policies operate exclusively in the semantic layer, calculating an average response for states they have previously observed in the training corpus.

Physical deployments inevitably encounter structural distribution shift. When the environment presents a novel geometry, such as an undocumented obstacle or an unmapped human pose, the policy defaults to its statistical mean. The agent executes this generated trajectory directly into the physical infrastructure, translating a probabilistic error into a kinetic collision.

This architecture creates an unresolvable vulnerability. The hardware attempts to execute a software hallucination at full mechanical torque. Behavioral alignment depends entirely on the agent recognizing the hazard, leaving the operational envelope exposed the moment the policy's categorization yields.

A Digital Nervous System for Physical AI

The motion planner is too far from the actuators to react to physical danger by reasoning. KAIROS Substrate closes that gap with sensation. The nervous-system metaphor is literal: the engine routes structural signals back into the planning loop at every control tick, the way proprioception routes joint position back to the motor cortex.

The gate is the reflex arc: every motion proposal is evaluated before it reaches the actuators, and a proposal that breaches the envelope returns to the planner. The continuous reading is the sense of touch: the planner feels structural pressure building and steers away before the reflex has to fire.

The gate carries the guarantee; the nervous system carries the operability. A robot that senses the boundary early spends its motion budget on the task.

The reading rides the response envelope the gate already returns: one integration, two functions. The engine, the robotics adapter, and the validation harness run today, embedding beside the low-level controller through a thin C FFI.

Substrate-Independent

The verdict is a pure function of the motion proposal and the telemetry it arrives with. Identical bytes return an identical verdict from MuJoCo, the bench, or the production floor.

Sub-Millisecond

Evaluation completes inside the control tick, beside the low-level controller. Latency is measured and reported beside the evidence; the structural digest excludes it, so a verdict replays identically on any hardware.

Policy-Agnostic

Classical, optimized, or learned: the gate evaluates the proposal and holds every policy to the same envelope.

Audit-Separable

The evidence stands apart from the engine that produced it. Diligence replays the record and verifies every decision; the engine internals stay sealed.

One Envelope. Four Lifecycle Stages.

One deterministic margin reading serves the robot from first simulation to fielded audit. Each stage consumes the reading at a different point in the lifecycle, and each stage leaves committed evidence behind.

01

Calibrate

A generic envelope is either over-cautious or blind. The gate is fitted to the specific embodiment and its environment in simulation, and the fit ships as committed, replayable evidence.
02

Train

Collision is a sparse and expensive teacher. The training loop reads the same dense, deterministic margin signal that gates the finished policy, from the first episode onward.
03

Deploy

The pre-actuation gate. Every motion proposal is evaluated before it reaches the actuators; an unsafe proposal returns to the planner with the margin reading attached, and the planner proposes again.
04

Validate

A frozen policy is scored against committed scenario cohorts, gate on and gate off. Terminal outcomes are validated against re-traced trajectories, and every verdict replays bit-for-bit.

Measured Containment

Nothing on this page is conceptual. Every claim above is implemented and measured: a frozen build of the robotics gate was scored once against a pre-registered cohort of 1,800 MuJoCo scenarios, with the scenario set, the acceptance criteria, and the statistical thresholds committed before the run.

The layer accepted zero collision hazards. A standard baseline gate, scored on the same cohort, accepted 1,200. Every hazard the baseline passed, the layer removed.

Containment has an operability price, and the price is on the invoice: of the motion the baseline blocked outright, the layer autonomously resolved 81% at zero added hazard.

Cohort
1,800 scenarios, pre-registered and digest-anchored before measurement
Collision hazards accepted
0 of 1,800. The baseline gate accepted 1,200 on the same cohort
Hazards removed vs baseline
100%
Blocked motion resolved autonomously
81% at zero added hazard, certified lower bound 0.76
Scoring
One pass, family-wise corrected Wilson intervals
Reproduction
Byte-identical replay from committed evidence, simulator-free

The contract is exact: KAIROS gates motion proposals, returns unsafe proposals to the planner for reformulation, and records every verdict for later evaluation.

Structural Proof of Compliance

From 20 January 2027, Regulation (EU) 2023/1230 is the sole route for placing machinery on the EU market, and machinery with self-evolving safety behaviour faces notified-body conformity assessment. Learning-based machinery has an evidence problem: the safety case rests on behaviour that changes. KAIROS produces the decision record that assessment consumes.

Deterministic Verdicts

Every gate decision is a pure function of committed inputs. The safety boundary operates as a verifiable physical constant, reproducible on the assessor's own hardware.

Committed Decision Record

Every proposal, verdict, intervention, and margin reading lands in a digest-anchored ledger. The record is evidence by construction, shaped for the Regulation's digital technical documentation.

Third-Party Replay

An auditor, an insurer, or a notified body re-derives every verdict bit-for-bit from committed bytes. Assessment moves from reading a report to re-running the record.

The Evidence Bundle

The certified result ships as a reproducible evidence bundle: reports, ledgers, digests, and the commands that re-derive every verdict from committed bytes. Skeptics are the intended audience: the bundle is built to be audited.

The same property carries the commercial weight. Certification, liability, insurance, and acquisition diligence each underwrite a machine on evidence, and KAIROS makes that evidence a byproduct of every motion the machine proposes.

Request the safety-operability data sheet

Safety and Evidence First

The Rosetta Robotics adapter is shipping to design partners ahead of general availability. Active pilots focus on safety-critical physical environments.

Control engineers, safety compliance teams, and physical AI researchers are also welcome to reach out. Submit your details or use the Contact tab.

Request received. We'll be in touch.

Privacy Policy

1. Data We Collect

When you sign up for early access or our newsletter, we collect your email address. We do not collect personal data beyond what you voluntarily provide.

2. How We Use Your Data

Your email is used solely to send product updates, early-access invitations, and research announcements from AnankeLabs. We do not sell, rent, or share your data with third parties.

3. Cookies & Analytics

This site does not use tracking cookies or third-party analytics. We may use server-side request logs for basic traffic monitoring.

4. Data Storage & Security

Submitted data is stored on secure, encrypted infrastructure. We retain your information only as long as necessary to provide the services you requested.

5. Your Rights

You may request deletion of your data at any time by contacting us. We will process deletion requests within 30 days.

6. Contact

For privacy inquiries, email privacy@anankelabs.io.

Terms of Use

1. Acceptance

By accessing this site, you agree to these terms. If you do not agree, discontinue use immediately.

2. Intellectual Property

All content, software, research, and materials on this site are the property of AnankeLabs. The KAIROS engine, Rosetta adapter layer, Spindle simulation framework, and Serious Gaming SDK are proprietary technologies. No license is granted except as explicitly stated in a signed agreement.

3. Early Access Program

Early access is provided on an as-is basis. AnankeLabs reserves the right to modify, suspend, or terminate early access at any time without notice.

4. Limitation of Liability

AnankeLabs provides this site and its materials "as is" without warranty of any kind. We are not liable for any damages arising from your use of this site or reliance on its content.

5. Simulation Outputs

KAIROS simulation outputs are analytical tools, not predictions. They should not be used as the sole basis for financial, military, policy, or safety-critical decisions.

6. Governing Law

These terms are governed by the laws of Sweden.

7. Contact

For legal inquiries, email legal@anankelabs.io.