Research roadmap
This page is the working plan for the book. It lists the open threads and the next step on each, and keeps a dated record of findings and decisions. It is updated at the end of every working session. The book itself is not planned in advance; chapters grow out of the findings recorded here.
How we work
- Session protocol. Each session starts by reading this page and ends by updating it: thread statuses, findings, decisions, and a changelog entry if anything was published.
- From finding to page. A question becomes a thread here. Exploration happens in scratch code. A result that holds up gets a function in
mibookand a test intests/, then a dated research note (status exploratory). The note is folded into a chapter once its result is checked, its proofs are labelled, and it connects to the existing material. The note stays as the dated record and links to the chapter. - Checked numbers. Every number in the text is computed by
mibookand pinned by a test. Figures are code cells that callmibook.figures. - Proofs. An incomplete proof is titled “Proof sketch” and states what it assumes.
- Citations. Each citation is marked full text or secondary in the sources appendix.
- Prose. All text (chapters, notes, this page, the changelog and commit messages) is written and edited with the humanizer skill: no staged contrasts, closers or inflated claims, no em dashes in prose, plain verbs.
- Releases. Dated tags (
v0.1,v0.2, …) with a changelog entry. Before a result goes to a venue, check that venue’s preprint policy. - Branches. Book content goes straight to
main, which deploys the site. Changes to other repositories, such as the lab website, go through pull requests. This is a default and can change.
Threads
Order is chosen together at the start of a session. Status is one of active, queued, parked or done.
T1. Foundations of the energy metric. Done, except for one open item.
- Where: 2 The energy metric, 7 Why energy.
- Open: confirm the citations marked secondary against the sources. The PDFs that could not be fetched automatically are Li, Li and Fritzen (2007); Moore (1981); Kubo (1966); Lee, Wensing and Park (2020); Zahedi and Ay (2013); Cover and Thomas, chapters 9 and 12; Kammer (1991); Geveci (1984), or Baker (1976) or Dupont (1973).
T2. Retained versus routed information. Queued.
- Question: what is the right definition of routing efficiency (sensor bits divided by full-state bits under Langevin noise), and what are its properties? Candidates to prove: bounds, invariance under relabeling the sensor, behaviour under composition of mechanisms.
- Next step: state the definition, prove the bound \(0\le\eta\le1\) and the invariances, and compute \(\eta\) for the joint encoder as well as the motor encoder.
T3. The Langevin body as the default sensor model. Queued.
- Question: can the discrete-time constant in Proposition 8.2 be proved, not only checked?
- Next step: expand \(B_d^\top Q_d^{-1}B_d\) in \(\Delta t\), including the position-block terms, and show the leading coefficient is \(J D^+J^\top/(2\sigma_T^2)\).
T4. Predictive validation on the Go2 designs. Parked: needs data.
- Question: which of retention, routing efficiency and \(I(W;X\mid A)\) best predicts closed-loop tracking error across designs and seeds?
- Depends on: the RL training pipeline and trained policies from the quadruped experiments, which are not in either repository.
T5. The Langevin body in the MuJoCo leg. Queued.
- Question: does the analytic Langevin result (flat full-state bits, motor bits climbing with stiffness) hold in the nonlinear simulated leg?
- Next step: inject damping-matched noise forces in the simulation and estimate with InfoNCE. Needs the
simextra (torch and warp), ideally on a GPU.
T6. Floating bases. Queued.
- Question: how should base positions that store no energy be treated?
- Options: keep velocities only; regularize with \(\omega_{\rm ref}^2H\) and sweep \(\omega_{\rm ref}\); use gravity and contact stiffness at the operating point.
T7. Nonlinear and multistable bodies. Queued.
- Question: what does the framework say about mechanical memory, where the noise law is the Gibbs law on a non-quadratic potential?
- Next step: a bistable element with an energy barrier; bits stored against barrier height and noise energy.
T8. Latency. Queued.
- Question: when does routed information arrive at the sensor, and how does that depend on the transmission? Extends the temporal Shapley analysis.
T9. Relevance. Queued.
- Question: which measures capture information relevant to the controller or the task, such as \(I(W;X\mid A)\) or control information?
T10. Literature. Queued.
- Confirm the secondary citations (see T1), then survey morphological computation, physical reservoir computing and the thermodynamics of computation.
Findings log
Newest first. Each entry gives the finding, the test that pins it, and where it is written up.
- 2026-11-21. Energy is the unique noise metric under five requirements, and invariance under free motion is equivalent to “free motion never creates information” (
test_uniqueness_step_four_identity,test_free_motion_creates_information_only_without_energy_noise; 7 Why energy). - 2026-11-21. The frequency-weighted family \(W_p\) converges in the stiff limit exactly when \(p<2\); only action weighting (\(p=-1\)) also passes mesh refinement (
test_frequency_weighted_stiff_limit_threshold_is_p_equals_two,test_frequency_weighted_mesh_behaviour; The alternatives compared). - 2026-11-21. Data-driven whitening gives every body the same bits (
test_data_driven_whitening_erases_the_body; Proposition 7.2). - 2026-11-21. Langevin body: with the full state observed, every design gives the same bits, set by the contact damping \(JD^+J^\top=1/b_i\). The rigid transmission routes 99.9% of that to its motor encoder; the SEA climbs from 6% to 99.5% with stiffness. Readout-only noise misranks resonant SEAs by up to 5.9 bits (
test_langevin_full_state_information_depends_only_on_contact_damping,test_langevin_routing_reproduces_the_proprioceptive_argument; 8 A thermodynamically consistent body: Langevin noise). - 2026-11-20. Coordinate invariance, the port-Hamiltonian structure, the stiff-SEA limit with explicit rates, the 8.4716-bit duplicate-readout artifact, SNR asymptotics, consistent Shapley noise, the closed form for \(\mathrm{MC}_W\), mesh independence and the floating-base condition (the remaining tests in
tests/test_energy_whitening.py; Part I).
Decisions log
- 2026-11-21. The work becomes a live textbook on GitHub Pages, published as results are checked; important results go to venues later. The book is not planned in advance.
- 2026-11-21. Quarto for authoring (one source for the website and the PDF; numbered theorems with cross-references; figures from code).
- 2026-11-21. A dedicated repository,
CyPhiLab/mechintelligence, owned by the lab. It depends on the analysis library inzpatty/quadruped-mech-info, pinned to commit2b25bca, so the book’s numbers stay reproducible. - 2026-11-21. The book is served at
mechintelligence.cyphilab.comand linked from the lab website’s navigation bar (CyPhiLab/CyPhiLab.github.io pull request 3). - 2026-11-21. Two-tier publishing: dated research notes, and chapters that absorb them once stable.
- 2026-11-21. All prose goes through the humanizer skill.
- 2026-11-21. Branch workflow: direct pushes to
mainfor book content; pull requests for other repositories. - 2026-11-21. The book is live at
mechintelligence.cyphilab.com, taggedv0.1. The Publish workflow installs Quarto and TinyTeX withquarto-dev/quarto-actions/setup, becausequarto install tinytexhit GitHub’s anonymous API rate limit on the shared runners.
Paper candidates
- The energy metric is unique, and free motion cannot create information under it. Theorem 7.1 and Proposition 7.1, with the comparison of alternatives. Possible venues: a robotics venue (RSS, ICRA, IEEE Transactions on Robotics) or a physics or information-theory venue.
- Morphology routes information; the contact sets the ceiling. Proposition 8.2 and the routing results, ideally after threads T2, T3 and T5.