Palm Garden AI develops Coherence Guard relational decision layer for human-facing robots - The Robot Report

As so-called general-purpose robots and humanoids continue to evolve, so is the software stack to enable them to conduct useful tasks around people. Palm Garden AI is developing Coherence Guard, which it described as a “platform-agnostic relational decision layer for human-facing robots.”

“The aim is not to replace perception, motion planning, reinforcement learning, or existing robot control stacks,” said Joachim Scheuerer, CEO of Palm Garden AI. “Rather, it functions as an additional pre-action evaluation layer: Before a robot executes an action, the layer can evaluate whether the action is relationally coherent in a real human environment.”

“This includes signals such as timing, proximity, boundary requests, emotional tone, trust preservation, respectful withdrawal, and the difference between technically possible action and socially appropriate action,” he added. “As humanoids move toward hospitality, care, retail, education, guidance, and domestic environments, we believe this may become a necessary infrastructure category.”

Palm Garden AI, which has offices in Germany and Thailand, has built its ANATTA 9 behavior infrastructure on the Tetrahedron Coherence Operating System (TCOS). The company said Coherence Guard is designed to sit above or beside existing robot control, SDK/API, ROS 2, planning, or world-model systems.

While physical world models help AI systems understand objects, space, and movement, Palm Garden said its Relational Infrastructure Framework (RIF) adds an understanding of roles, intentions, vulnerabilities, and possible future consequences.

The technology can evaluate human expressions and guide coherent actions, such as withdrawing if a person indicates discomfort. The RIF Relational Infrastructure Framework is now available upon request from Palm Garden.

Palm Garden AI adds a layer to robot understanding

Scheuerer replied to the following questions from The Robot Report:

How did you identify the need or gap in current service robot capabilities?

Scheuerer: We saw the gap from two directions. First, many current service robots are already becoming capable in navigation, speech, perception, task execution and expressive interaction.

But in real human environments, the difficult moment is often not the task itself — it is the relational decision around the task: when to approach, when to pause, when to withdraw, how much to explain, how to handle hesitation, discomfort, confusion or changing boundaries.

Second, our work at Palm Garden Retreat in Thailand exposed us to many real-world human interaction situations: arrival, orientation, guidance, silence, vulnerability, trust-building, misunderstanding and respectful withdrawal. These are situations where a technically correct action can still feel wrong if timing, distance, tone or context are not coherent.

Coherence Guard was developed to address this missing layer — not replacing robot control, but evaluating whether a proposed action is relationally appropriate before or during execution.

Do you have base behaviors based on your observations of human interactions?

Scheuerer: Yes. We have developed a set of base behavior patterns from three years of structured observation, retreat practice and human interaction training. These include greeting and orientation, supportive presence, non-intrusive assistance, respectful withdrawal, escalation when uncertainty is high, and coherence-preserving explanation.

One simple benchmark is “respectful withdrawal.” If a person shows discomfort or asks for space, the robot should not simply continue the task. It should pause, acknowledge the signal, increase distance if appropriate, reduce expressive intensity, and return to a neutral or available state. We see this as a core service-robot behavior, especially for hospitality, eldercare, guidance, and domestic environments.

Does your company have experts in human-robot interaction (HRI)? Are there precedents in other technologies?

Scheuerer: Palm Garden AI is not a traditional academic HRI lab. Our core expertise comes from long-term work in human interaction, psychotherapy-related software, retreat facilitation, relational training, architecture of human environments, and AI behavior design. We are now applying this background to human-robot interaction through a dedicated robotics layer.

There are precedents in other technologies. Aviation and automotive systems use safety monitors and override logic; collaborative robotics uses safety envelopes; AI systems increasingly use guardrails and policy layers; and autonomous systems often separate task planning from safety or governance checks.

Coherence Guard follows a similar principle but applies it specifically to relational coherence in human-facing robot behavior.

---

Save the date for RoboBusiness 2026

---

Coherence Guard to complement existing safety systems

How will your system work with evolving safety standards for robots — humanoids in particular?

Scheuerer: We see Coherence Guard as complementary to formal safety systems, not as a replacement for them. Certified robot safety must remain at the hardware, control, emergency-stop, collision-avoidance and risk-assessment levels.

Our layer sits above or beside those systems. It evaluates candidate actions from a relational and contextual perspective: Should the robot continue, pause, explain, ask for confirmation, reduce proximity, or withdraw?

As humanoid standards evolve, we expect such layers to become more important because humanoids operate closer to people and are often socially interpreted by users. Coherence Guard is designed to support auditability, logging, scenario testing and configurable thresholds so it can adapt to different compliance environments.

Where does Coherence Guard run — on the edge device, on premises, or in the cloud?

Scheuerer: The architecture is designed to be flexible. For latency-sensitive or privacy-sensitive situations, Coherence Guard can run on the edge device or on premises. For simulation, analytics, configuration, model improvement or fleet-level learning, cloud components can be used.

Our preferred deployment model for human-facing robots is local-first. The immediate relational decision should not depend on cloud latency. Cloud can support updates, scenario libraries, logs and non-real-time analysis, but the real-time coherence check should be close to the robot.

Software is available to hardware partners

Are you offering it through a software-as-a-service (SaaS) model? How open is the software?

Scheuerer: We are currently preparing the commercial model. The likely structure is a licensed software layer with optional SaaS components for configuration, simulation support, analytics and updates.

The core IP is patent-pending, so it will not be fully open-source at this stage. However, we want the integration interfaces to be as open and platform-agnostic as possible. We are designing around ROS 2, SDK/API compatibility, simulation-first workflows and adapter layers, so robot manufacturers do not need to replace their existing stack.

With the simulation-first pathways, how do you ensure that you have the right data and conclusions?

Scheuerer: We are careful not to treat simulation as final proof. Simulation is the first filter. It allows us to test defined scenarios, compare candidate behaviors, log decision traces, and identify failure modes before using real hardware.

The pathway is staged. First, logic simulation, then ROS 2 or platform simulation using URDF or SDK interfaces, then limited real-robot pilots. The conclusions from simulation are framed as compatibility and behavioral hypotheses, not final claims.

The key is to define narrow, observable benchmarks — for example, approach distance, pause timing, withdrawal behavior, explanation level and escalation triggers — and then validate them with real human feedback.

Are you already working with robotics hardware and software providers?

Scheuerer: We are in active technical and partnership evaluation with several robotics providers. With Robotera, we have already had a technical call and are moving through an NDA and simulation-first compatibility pathway.

With Hanson Robotics, the compatibility path has been discussed, and we are preparing the next phase under NDA/addendum. We have also evaluated interface compatibility with other platforms, including ROS 2/SDK-based humanoid systems, and we are mapping possible connections to NVIDIA Isaac/GR00T-style simulation and middleware environments.

At this stage, we describe these as technical evaluations and pilot discussions rather than completed commercial deployments.

As you work to get patent approval, what are your next steps?

Scheuerer: Our next steps are:

1. Finalize the patent-pending technical framing around TCOS, FIE, and Coherence Guard
2. Complete Phase 0 compatibility reviews with selected robot platforms
3. Build and document simulation-first benchmarks for human-facing service scenarios
4. Run a limited pilot focused on greeting, guidance, explanation and respectful withdrawal
5. Prepare a clearer technical package for robotics companies: architecture, integration points, benchmark scenarios and commercial licensing options

Our goal is not to create another robot body or another conversational AI system. Our goal is to provide a relational decision layer that helps service robots behave more coherently, safely, and respectfully in real human environments.