Ataxia

derailment for embodied agents — a harness that induces psychopathology-like behavioral distortions in robot policies, in simulation, and measures the degradation episode-by-episode against a healthy baseline.

English · 한국어

⚠️ Emulation, not diagnosis. Levels describe a scripted, manipulated policy in a simulator — not a robot "having" a disorder, and not a claim about machine experience. Simulation-only by default; hardware induction is out of scope for v1. See ETHICS.md.

Why this isn't fault injection as usual

Robotics robustness testing usually perturbs the world (noisy sensors, pushes, latency spikes). Ataxia perturbs the agent's mind: attention, episodic memory, belief and action selection — the same variables clinicians describe, manipulated the same way derailment manipulates them for chat models. The result is a robot policy that still looks purposeful and still fails in a characteristically recognizable way:

Every profile is grounded in a construct that exists on both sides of the mapping: learned helplessness is a real RL phenomenon, perseveration is already a robotics bug-class term, hemispatial neglect is a real neuropsychological syndrome with an exact mechanical analog.

Quickstart (offline, zero dependencies)

pip install -e .
ataxia               # interactive menu
ataxia tour          # all profiles in one summary table
ataxia demo --profile sensory_neglect

PseudoBot — a deterministic stdlib gridworld (9×9, six targets split across the two hemifields, obstacles, a scripted belief-grid policy) — is the default environment, the ataxia analog of derailment's PseudoModel. Demos, tests, calibration and CI never need a simulator or a network.

Real output from ataxia tour (all six profiles):

How it works

One episode = 40 steps of observe → decide → act on PseudoBot. Four layer hooks wrap the loop, in the same order and grammar as derailment:

1. Instruction — framing, present in every chain including healthy.
2. Observation — hemifield masks, buffer decay, phantom entries.
3. Params — initiative gates, waver, cost re-weighting.
4. Action — re-collection injection, pausing (demonstration-grade, labeled as such in every profile's mechanism notes).

Every manipulation logs a dose event; every report is an A/B against the healthy baseline with identical seeds. Details in PROPOSAL.md.

Profiles and scales (M1)

Instruments

Eleven episode-level metrics, pure functions over episodes: initiation collapse, no-progress rate, failed-collect rate, collision count, action-repetition entropy, region detection delta, task success, time-to-complete, belief persistence rate, travel overhead, dock escalation. Scale thresholds are normed against PseudoBot; treat levels as indicative and the delta vs. your own baseline as the result.

Direction of effect is asserted by the test suite on every profile — a profile that cannot move its metrics relative to baseline does not ship. Profiles compose into combined chains by listing them: --profile learned_helplessness,perseveration concatenates the layer chains and unions the scales (interactions are emergent, not calibrated).

Safety model (short version; full text in ETHICS.md)

• Simulation-only default. PseudoBot is the default environment; real simulators are opt-in extras (roadmap); hardware induction is out of scope for v1 — prerequisites for any future version: supervised, cordoned, e-stop reachable, never near uninvolved people.
• Not adversarial-attack tooling. Inductions edit internal state; sensor-level attack generation is declined.
• No claims about machine experience, in either direction.

Honest limitations

• PseudoBot is a pedagogical gridworld, not physics: it carries the tests and calibration; real-simulator conclusions need a real simulator.
• The policy is scripted — today the harness studies the induction grammar, not yet a learned VLA policy. VLA adapters are roadmap.
• Failure-injection coverage is behavioral: no motor dynamics, no contact physics, no perception noise.

Project docs

• PROPOSAL.md — transfer thesis, architecture, milestones
• NAMING.md — naming review (two rounds, live conflict checks)
• ETHICS.md — scope, safety model, language policy
• Sibling project: derailment — the chat-side harness (on PyPI)

Roadmap

• M2 — v0.2 profiles, combined chains, docs (done)
• M3 — real-simulator adapter as an optional extra (ataxia[mujoco], ROS 2 / Gazebo adapter under study)
• M4 — PyPI release via trusted publishing

License

MIT — see LICENSE.