Equilibrium is a domain-specific language for supervisory interaction. Its reason to exist is the common representation of physical dimensions, sensor freshness/uncertainty, operator requests, independent effect authority and recorded outcomes. It does not replace motor control or safety controllers.
| Module | Executable responsibility |
|---|---|
| syntax.py, parser.py | tokens, positions, immutable source tree, precedence parser |
| units.py, expressions.py | dimensional types, decimal arithmetic, finite expressions |
| compiler.py | semantic checks, reachable states, bound summaries, checked artifact |
| sensors.py, robust.py | per-channel validity, quorum intervals, conservative predicates |
| runtime.py | snapshot plans, state transitions, bounded adaptation, effect mediation |
| authority.py | external scoped grants, exact-action reviews, expiry, revocation and reservation |
| evidence.py | durable commit-before-effect journal and independent integrity verification |
| adapters.py | host protocol and kinematic virtual plant |
| transport.py | authenticated scoped UDP sensor packets |
| simulation.py | scenario execution, replay, boundary faults and finite exploration |
| cli.py, report.py | operational commands and offline evidence display |
Compilation produces a revision-bound checked source tree with inferred dimensions and static bound summaries. Loading serialized IR recompiles and requires exact equality. The runtime interprets that checked tree; it does not generate Python or machine code from source. Its finite language excludes unbounded loops and dynamic calls, making each tick's source evaluation bounded in structure. That does not imply a bounded wall-clock execution time, because storage and host callbacks can block.
The external grant and sensor inputs are independent of source declarations. An operator intent input indicates a request. It is not permission. A review covers the exact proposed setpoints, source revision, session, state, time and tick. The source cannot manufacture that review. Adaptation cannot enlarge its declared parameter domain or actuator range.
The included demos cover assistive interaction, temperature mesh setpoints and paired actuators. The same primitives can express haptic envelopes, operator shared control and instrumented experimental meshes. Such applications still need actual validated sensors, adapters, dynamics and independent hardware limits. Neural inputs can arrive as host-fed channels, but this language does not infer mental states or decode EEG itself.
These are future engineering directions, not implemented features or promises:
- A native backend with checked-IR semantic equivalence tests and measured execution deadlines, while leaving real-time inner loops in suitable hosts.
- Coordinate-frame/vector types and transforms that reject frame mismatches.
- Tested ROS 2 integration and one instrumented hardware demonstration with independently enforced limits and measured stop behavior.
- Durable revocation, authenticated identity integration and signed checkpoint anchoring; a hardened journal with incremental validation and measured scaling.
- Temporal verification against explicit environment models, including sensor delay and actuator dynamics, with assumptions exported alongside results.
- Model-informed experimental soft-device or haptic systems, validated against real measurements rather than inheriting simulation conclusions.
No novelty/patent clearance, industry adoption or physical stability claim is made. A new syntax alone is not evidence of an advantage over other robotics tooling. Evaluate the implemented checks and recorded behavior directly.