⚠ See the matched-control correction (Part 49) — the matched-control result revises the “abstraction beats” cells in the unified table: vs a budget-matched PPO the abstraction wins sample-efficiency only, not the asymptotic ceiling.

We ran both algorithms — tdmpc-glass and vanilla tdmpc2 — across the full DMControl task suite at 1M env-steps, same box, same harness, same seed, so the comparison is apples-to-apples. The result refines the project’s long-standing “glass ≈ TD-MPC2” intuition into something more precise.

Setup

Deterministic rolling sweep (full_sweep_runner.sh) on a 4×3060 box: 16 DMC tasks × 2 algos, 1M steps each, --no_plot, returns read from exp/benchmark/*.csv (last logged eval, never a training-loss proxy). Primary numbers are seed-1 (both algos on the same box → matched seed). A second seed (reverse-order, separate box) is used only to cross-check anomalies. QuadrupedRun is omitted (not present in this mujoco_playground build).

Results (seed-1, 1M steps, return)

task tdmpc-glass TD-MPC2 verdict
AcrobotSwingup 312 383 TD-MPC2
BallInCup 973 976 tie
CartpoleBalance 982 988 tie
CartpoleSwingup 857 852 tie
CheetahRun 652 706 TD-MPC2
FingerSpin 963 977 tie
FingerTurnEasy 977 773* glass*
FingerTurnHard 586 970 TD-MPC2 ≫
HopperHop 0.17 234 TD-MPC2 (glass collapses)
HopperStand 355 872 TD-MPC2 ≫
PendulumSwingup 743 827 TD-MPC2 (both solve)
ReacherEasy 779 983 TD-MPC2
ReacherHard 973 971 tie
WalkerRun 665 ~625 (conv.) ~tie
WalkerStand 976 ~943 (conv.) ~tie
WalkerWalk 959 (converging) pending

*FingerTurnEasy: glass 977 vs TD-MPC2 seed-1 773 — but TD-MPC2 seed-2 = 978, so the seed-1 dip is seed variance, not a robust glass win. Read it as a tie. (Walker-trio TD-MPC2 still climbing at writeup; seed-2 TD-MPC2 = 651/972/975, and glass already sits at 665/976/959, so these are ties.)

⚠️ UPDATE (seed-2 in → n=2; seed-3 running): the single-seed story was too harsh on glass

The original table below was seed-1 only. A second seed substantially revises it — and the corrected story is about stability, not a flat capability gap:

  • HopperHop glass: seed-1 0.17 → seed-2 405.6. The “catastrophic failure” was a single-seed collapse, not a consistent failure. (TD-MPC2 solves it both seeds: 234 / 350.)
  • HopperStand glass: 355 / 14.7 — another collapsing seed. TD-MPC2 stays solid (872 / 858).
  • Several seed-1 “TD-MPC2 wins” shrink to ~ties at n=2: CheetahRun (glass 652/649 vs 706/605), FingerSpin (963/968 vs 977/925), ReacherEasy (779/979 vs 983/975).
  • Robust TD-MPC2 advantages that survive n=2: FingerTurnHard (970 vs ~674), HopperStand (865 vs ~185, glass collapse), AcrobotSwingup (378 vs ~263, glass also high-variance 312/285/191).

Revised verdict: tdmpc-glass is competitive-but-unstable — high seed-variance with occasional hard collapses (HopperHop, HopperStand), whereas TD-MPC2 is consistent. Most apparent per-task gaps are driven by glass’s collapsing seeds, not a uniform deficit. Pendulum’s earlier 0.0 stays debunked (both tdmpc2 seeds 827 / 725). Live mean ± 95% CI (n filling to 3) + wall-clock + per-task curves: the results page (auto-updating). The seed-1 tables below are kept for provenance; trust the live page’s multi-seed numbers over them.

All methods, one table (PPO · TD-MPC2 · tdmpc-glass · abstraction · TAMP)

Putting every method the campaign has on DMC into one place. Coverage is uneven and that is the honest point: PPO was run only on the handful of DMC tasks the campaign needed (and at 50–285M steps, vs 1M for the world models — not sample-efficiency comparable, it’s “PPO’s best at a far larger budget”); the abstraction (analytic controller + in-loop residual) exists only for the two swing-ups where we hand-built a controller; and TAMP is an OpenCabinet manipulation controller — there is no TAMP for DMC locomotion, so that column is N/A by construction (it is not a DMC method).

task PPO (50–285M) TD-MPC2 (1M) tdmpc-glass (1M) abstraction (ctrl+residual) TAMP
AcrobotSwingup 268 383 312 71 (backfires, Part 47) —
BallInCup — 976 973 — —
CartpoleBalance — 988 982 — —
CartpoleSwingup — 852 857 — —
CheetahRun 928 706 652 — —
FingerSpin — 977 963 — —
FingerTurnEasy — 773/978† 977 — —
FingerTurnHard 968 970 586 — —
HopperHop 33 234 0.17 — —
HopperStand — 872 355 — —
PendulumSwingup — 827 743 835 (ties/beats, Part 45) —
ReacherEasy — 983 779 — —
ReacherHard — 971 973 — —
WalkerRun — ~625 665 — —
WalkerStand — ~943 976 — —
WalkerWalk 970 ~975 959 — —

(†TD-MPC2 FingerTurnEasy: seed-1 773 / seed-2 978 — a seed dip, treat as ~tie. TAMP is reported on its own task in Part 43: residual-on-TAMP 0.98 success on PandaOpenCabinet vs PPO 0.98 — a manipulation result, not a DMC return.)

What the unified view shows:

  • The abstraction is bimodal, exactly per its prior quality: on PendulumSwingup it (835) beats both TD-MPC2 (827) and is the best method on that task; on AcrobotSwingup it (71) is the worst — same recipe, inverse outcome (Parts 45/47).
  • HopperHop is hard for everyone except TD-MPC2: PPO 33, glass 0.17, TD-MPC2 234. It is the exploration- bottlenecked task where model-based planning earns its keep — and where glass’s failure is most visible.
  • Where PPO data exists, TD-MPC2 matches or beats it at ~100× fewer steps (FingerTurnHard 970 vs 968, WalkerWalk ~tie, Acrobot 383 vs 268) except CheetahRun (PPO 928 at 187M » TD-MPC2 706 at 1M) — consistent with the Part 18–21 finding that TD-MPC2 wins per-step sample-efficiency while PPO can win final return at a huge budget.
  • abstraction/TAMP are not general DMC learners — they are task-structured methods that win exactly when their prior fits (Pendulum, OpenCabinet) and are absent or harmful otherwise. That is the campaign’s thesis in one table.

The refined verdict

TD-MPC2 ≥ tdmpc-glass across DMC, by task difficulty:

  • Ties on easy/saturated tasks — BallInCup, both Cartpoles, ReacherHard, FingerSpin, Walker-stand/walk: both hit the ~950–990 ceiling, indistinguishable.
  • TD-MPC2 wins the harder exploration/locomotion tasks — FingerTurnHard (970 vs 586), HopperStand (872 vs 355), ReacherEasy (983 vs 779), CheetahRun (706 vs 652), AcrobotSwingup (383 vs 312). The structural-entropy/graph machinery in glass does not help — and modestly hurts — where the task needs more raw value/dynamics fidelity.
  • One catastrophic glass failure: HopperHop = 0.17 while TD-MPC2 reaches 234 (and 255 on seed-2). HopperHop is the campaign’s known exploration-bottlenecked task (Part 18/21); glass simply fails to get off the ground on it.

So the earlier “glass ≈ TD-MPC2” (Part 2) was true on the subset of tasks tested then (mostly saturated ones). Across the full suite it becomes: glass matches TD-MPC2 where the task is easy, trails where it is hard, and has a failure mode (HopperHop) TD-MPC2 does not. An honest negative for the glass variant as a general DMC learner — its value was always the interpretability/structure angle (Parts 24–35), not raw DMC return.

Verification-discipline note (a debunked “finding”)

A seed-2 TD-MPC2 run logged PendulumSwingup = 0.0 — which looked like “TD-MPC2 fails a trivial task.” It is not: the seed-1 run reached 827 (glass 743), so both algorithms solve Pendulum and the 0.0 was a single-seed divergence/NaN. Reported here as an outlier, not headlined as a result — exactly the kind of single-seed artifact this project has learned to cross-check before claiming (the HopperHop collapse, by contrast, is reported because it reproduces and the task is provably solvable by the other algo).

Caveats

Primary numbers are single-seed (seed-1), 1M steps; a second seed cross-checks the two anomalies only. Walker-trio TD-MPC2 was still converging at writeup (cited values are seed-2 + the climbing seed-1, both ~tie with glass). Returns are the last logged eval per run (exp/benchmark/{tdmpc-glass,tdmpc2}_<task>_fs_*.csv). HopperHop glass collapse is seed-1 here; the matched scorer/harness make the cross-algo gap (0.17 vs 234) far larger than seed noise. Prior: Part 2 (glass≈TD-MPC2 on subset), Parts 18–21 (TD-MPC2 vs PPO on DMC, HopperHop exploration).