Research threads

1 · Hydraulic backscatter proprioception

Active research · potentially patentable · not a baseline requirement

An electronic sensor at every muscle is one way to know limb state, and an expensive one. The alternative here: a root transducer injects a small pressure chirp into the network already carrying power, and each branch modulates the acoustic reflection it sends back. For branch i:

where Z_m,i is effective mechanical impedance, A_p,i maps mechanical motion into hydraulic impedance, Z_0,i is line characteristic impedance, L_i is path length, c_i is pressure-wave propagation velocity, and τ_i is round-trip delay.

Conceptually: path delay identifies which branch reflected the chirp; reflection magnitude and phase change with actuator position, compliance, loading and contact; pulse compression at the root separates overlapping returns; and local actuator electronics may become unnecessary for some proprioceptive channels. The hydraulic distribution network becomes both a power bus and a sensing bus — a closer functional analogue of a nervous and vascular system.

Unresolved

• Dispersion and branch reflections
• Valve-state coupling
• Flow noise and pump noise
• Temperature dependence
• Signal bandwidth
• Identifiability in dense manifolds
• Calibration drift
• Whether sufficient state is recoverable at acceptable excitation amplitude

This is a research track, not a solved feature, and it is deliberately not a blocker for the first working body.

2 · Calibration-aware embodiment

Active research · core design principle

In software agents, many hard problems arrive already discretised into “click”, “call an API”, “run a command”. In robotics, “pick up the object” expands into camera calibration, object segmentation, pose uncertainty, camera-to-world and robot-to-world transforms, trajectory generation, backlash and compliance, contact detection, friction, grasp closure, load verification, and recovery from slip or collision. Measurement construction is therefore treated as part of the task rather than invisible infrastructure.

Instead of a hidden assumption that “the camera is calibrated”, the system knows something closer to:

calibration:
  head_camera_extrinsic:
    estimate:    …
    covariance:  …
    observed_at: …
  left_gripper_tcp:
    estimate:    …
    covariance:  …
  valve_zero:
    zone_2_axis_4: …
  time_sync_error_ms: …

This permits the planner to ask whether an action is limited by uncertainty about the world or by uncertainty about the self — a distinction most stacks cannot represent at all.

3 · Event-sourced autonomy

Active research

Logging discards structure; this does not. Observations, motor commands, calibration changes, hypotheses, tool calls, safety events and model updates land in one typed causal history, parented, and that history is what both audit and training read from later.

Event
  event_id · timestamp · stream_id
  stream_type ∈ { sensory, motor, tool, calibration,
                  thought, judge, train, safety, network }
  source · payload_ref · summary
  parent_event_ids · causal_tags
  uncertainty · calibration_state · model_state_hash
  safety_state
  world_state_delta · belief_delta · self_model_delta
  training_candidates

The desired property is that a claim such as “the cup is now in the sink” can point back to the action, observations, contact events, pose estimates and verification frames that support it.

Persistent ledger, transient workers

One persistent global belief and task ledger forks bounded transient reasoning streams for local work — inspect workspace, identify an object, plan a grasp, calibrate a camera, diagnose a leak, check route feasibility. Workers return structured merge packets rather than raw monologues:

claim: …
evidence_event_ids: [...]
confidence: …
uncertainty_delta: …
belief_delta: …
proposed_next_actions: [...]
unresolved_questions: [...]
failure_modes: [...]
training_candidates: [...]

This is deliberately different from pretending an infinitely growing chat transcript is a durable mind.

4 · Causal agency as the objective

Active research

Intelligence, here, is the capacity to bring an internal transition model into alignment with the world's actual one through intervention. Observation on its own will not do it. Information gain appears as (C3), self-effect as (C4); two further quantities carry no closed form yet:

• Calibration gain — uncertainty reduction caused by an explicit calibration action.
• Affordance expansion — increase in the number or quality of actions and measurements made possible by learning something or constructing a reusable skill.
• Replay auditability — fraction of conclusions and policy updates reconstructible from recorded evidence.

These are not a completed objective function. They define a direction: reward the system for making reality more predictable through grounded intervention, not for producing plausible narration about what it would do.

5 · Morphology-conditioned policy

Software / research architecture

Nothing learned should hard-code one monolithic vector of joints and sensors. The body is naturally a typed morphology graph.

This supports component replacement, simulation generated from the same morphology definition the robot runs, model transfer across body revisions, graph-based state estimation and control, discovery of local interaction structure, and an explicit mapping between physical topology and learned topology. Learned dynamics can be conditioned on the graph rather than assuming every machine has exactly the same body.

6 · Design automation

Active software direction

The fabrication strategy becomes substantially more valuable if geometry is generated from a machine-readable system model rather than manually redrawn each revision. Associated work has included a generative CAD workflow and experimentation with a multiphysics simulator spanning mechanics and fluid behaviour. The high-value version is not “AI makes pretty CAD”; it is a compiler-like path from requirements to geometry and verification artefacts:

This closes another reality loop: CAD parameters, simulator parameters, controller parameters, calibration state and measured hardware should describe the same component identities rather than living in disconnected tools.

7 · Verification matrix

What we intend to publish against, layer by layer. This table is the standard we would like to be held to.

8 · Claims discipline

Every quantitative public claim carries one of: measured on prototype, simulated under stated assumptions, engineering target, historical exploratory estimate, research hypothesis.

Get the build log →