What is stopping
the controller?

More power changes the controller, too

Try the discrete simulation: the ready-to-play menu now includes NEAT and PPO 140° from rest · 120 Hz experiment entries. Both use doubled acceleration and count catches involving the rail. The top-bar Physics selector shows each demo’s recorded rate; select another rate to start a fresh temporary trial with the same weights and controller frequency.

Six distinct saved double-pendulum networks were tested with larger acceleration limits, releases from rest and physical knockdowns. The best-looking high-acceleration NEAT swing-up did not survive finer physics steps. None of these changes established a reliable swing-up without rail contact.

Original audit: would the catch survive a smaller timestep?

Both rods start 140° from upright: 50° below horizontal, at rest. Each cell counts runs holding every link above horizontal for the final 15 seconds of a 90-second rollout. Twenty fresh starts per cell. These are observed holds, before rejecting large constraint errors; every successful large-angle run below contacted the rail.

The separate PPO 25° actor often finds a rail-contact catch at its original limit. That is an interesting experimental behavior, with failures still present; it does not establish that higher acceleration is the missing ingredient. The NEAT reversal shows strong timestep sensitivity, although these tests alone do not isolate which part of the contact dynamics causes it.

120 Hz is not universally easier. In the same audit, the momentum actor released at 140° with zero velocity and its original acceleration held 0/20 at 120 Hz, 0/20 at 240 Hz and 17/20 at 480 Hz, again with rail contact. A smaller timestep can help one frozen controller and defeat another. These observations describe the discrete simulator and policy together.

Next experiment: keep observation units fixed, then train with a range of acceleration limits and recovery states. Measure when the actuator saturates, whether each push adds useful energy, and whether catch rate survives smaller timesteps. A failed frozen policy does not prove a state is physically unreachable. Energy and phase matter alongside actuation limits, as illustrated in MIT’s underactuated-control notes.

Hold, acquire, and verify are three separate gates

Angles are measured from upright. The learned records keep their full scene and held-out results. Their playback uses only the neural policy. The three references compute controls from a model; no neural weights are implied.

Did the topology change actually help?

On the same 40 fresh starts in the new 42° scene, the earlier NEAT policy passed 0/40 and its evolved continuation passed 40/40. Hidden nodes grew from10 to11; enabled connections changed from66 to51. The improvement is in the measured behavior, not simply a larger graph. The new policy also passed finer-step and independent nudge tests.

A successful trajectory is evidence, not the finish line

Trajectory optimization found full swing-up motions in the actual point-mass simulator at high gravity. Those first PBD plans failed when the physical timestep was refined while controller timing stayed fixed, so they remain experimental teachers. The rigid references passed this check after refinement of their trajectory design. This distinguishes an executable strategy from a numerically robust controller.

Map a frozen network’s release envelope

Each cell runs two opposite releases for 30 simulated seconds. Green means both keep every link above horizontal for the final 10 seconds, with no rail contact anywhere in the trial. This is a quick diagnostic, not a reliability estimate.

Ready. These experiments run locally using the production physics and network.

Columns are first-link angles from upright: 90° is horizontal; 140° is 50° below horizontal. Other links retain the scene’s linked initial pose. Actual magnitudes add a fixed 0.35° perturbation, shared by every row. This matrix tests releases from rest; the supplied-momentum case has its own experiment below.

Refinement preserves the controller’s sampling period and damping per second. The stronger-actuator row preserves the saved velocity-input normalization; it changes the plant, not the weights. Nudge tests add ±35 px/s to the tip at 10 and 20 seconds. No assistance, teleportation, automatic reset or training is used.

A moving start changes what is recoverable

Both double-pendulum rods begin about 105.5° from upright, below horizontal, with inward angular rates of about 5 and 1 rad/s. The older frozen PPO holder catches them using cart acceleration alone at g 8.5 on the original ±260 px rail. The initial velocities are consistent with the rod constraints; every subsequent frame runs the physical simulation. This is momentum-assisted swing-up, not recovery from rest or an animated state correction.

Each trial lasts 90 s, with all links continuously above horizontal for the final 30 s and no rail contact; control stays at 120 Hz. Starts vary both rod angles and rates, both mirror signs, and initial cart velocity. Passing this bounded family does not imply arbitrary disturbance recovery.

How long should I watch?

The ready controller strip shows measured simulation-time ranges. The PPO catch takes about 2.4 s, including a two-second confirmation within 60° of upright. Rigid-reference timings use their actual holding-phase test. A displayed catch allows motion; it is not a claim of motionless equilibrium. Manual pushes or edited parameters invalidate the undisturbed timing estimate.

A learned time estimator is practical with a sufficiently broad rollout dataset. It should predict a time distribution and the chance of never catching, keeping failed or unfinished runs as censored observations. For a small fixed catalog, measured ranges are easier to validate than another neural model. Neural time-to-event prediction

Predict motion, then learn which route to take

The older dodge CNN already sees the whole board. Its danger encoding, however, can alias bullets moving in different directions. The new controller supplies nine short physical forecasts and uses a CMA-ES-trained shared network to rank them. Forecasting is an explicit model; the 41 scoring parameters are learned. There is no fallback controller or knowledge of future bullets.

For each command, the forecast measures clearance over four time horizons, wall margin, centrality, velocity alignment and continuity with the previous command. It predicts currently visible bullets ballistically: b(t) = b(0) + vt. The learned score chooses a route; a high score is a preference, not a calibrated probability of survival.

The forecast makes looming collisions explicit before contact, reducing what this small network must discover. This works well for the current ballistic bullets, but a forecast can be wrong and nine constant commands do not cover every evasive maneuver. Newly spawned bullets are unknown until observed. A temporal CNN could learn a richer policy; it needs its own matched training experiment before claiming which method is better.

Mixed attacks at spawn-rate setting 3 (emissions vary by pattern); forty fresh seeds, paired terminal and continuous tests. The current CNN and dense policies were replayed, not retrained with matched compute. These results compare available controllers, not an architecture's ultimate capacity. At twice the spawn rate, the scorer survived 12 of 20 fresh trials.

Limiting the new forecast to the nearest eight bullets still gives 33/40 at this density. The evidence therefore supports usable motion information; it does not demonstrate that the wider field caused the improvement. The Network tab shows actual candidate scores, forecast paths and past frames. Those history thumbnails are diagnostic only. A temporal CNN is a sensible next controlled comparison, but it has not been trained here.

Write the first word, not just a later loop

The old signing model's first word averaged 126 px of pen error on fresh starts, even though it could trace later loops. Its refinement discarded that first loop and froze a saturated cart/joint backbone. A tighter tolerance circle could not repair that training contract by itself.

The new experiment begins with neutral commands, learns 13 sparse/shared tanh coefficients and scores the entire 40 s first word. The cart and telescope outputs are learned; the existing joint servo holds a neutral bend. The target curve and future-target cues are supplied. All pen motion comes from the physical simulation.

Matched starts, initialization, architecture and budget: two paired seeds, 100 generations, population 24, two 40 s rollouts per candidate. Fixed tolerance is 30 px; shrinking uses 160 → 120 → 90 → 60 → 45 → 30 px. Selection used separate development poses before these 40 unseen starts. Every test includes three full words; no warm-up is discarded. Ordinary browser Train creates a fresh dense network; the saved sparse/shared search ran offline.

The diagnosis is trainability, not demonstrated impossibility. Both curricula fix most of the old acquisition failure. The circle adjusts reward precision; initialization, action parameterization and including acquisition in the objective mattered together. The first B still has an approach stroke, and a loop's mean error does not bound every point. This spring/servo scene is calibrated at 120 Hz: 240 Hz remains close, while 480 Hz degrades, so numerical convergence remains unresolved.

A clean strike through a constrained course

The two ridges alter launch and landing angles; a movable bumper transfers momentum; the low beam limits high arcs. Beam contact is legal and produces a passive rebound. After the first real club impulse, the club freezes. The network cannot steer the airborne ball.

Separating overlaps previously injected spurious velocity, and cached velocities could disagree with contact impulses. The challenge now applies impulses only to inward contacts and keeps the simulation's velocity representations consistent. A second issue was optimization: weights near ±15,000 saturated tanh units, leaving ordinary small mutations ineffective. Rescaling those weights into a trainable range and continuing CMA-ES produced the new frozen model on the same geometry and motor limits.

Commands remain 120 Hz. The same 200 reserved initial-angle/hole pairs test every rate. Club tilt is 3–7° left, hole center 240–280 px, width 100 px. These are center-based slot sinks, not a detailed cup/rim model. An attempt normally lasts 10 s, with up to 4 s extra while a struck ball is still moving quickly. The inherited cart uses a bounded target-velocity servo. Ball and hole state are inputs; obstacle geometry is a fixed training layout. Original golf and putt retain their exact earlier behavior.

There is useful structure before training begins

Finite differences of the actual point-mass simulation give full local controllability rank near upright for both double and triple pendulums. The legacy observation maps, however, hide some instantaneous state directions. The angular mode exposes every link’s signed rate.

Measured at gravity 1000, direct acceleration, no helpers. Scaled constrained angular coordinates; central differences ε = 10⁻³, 10⁻⁴, 10⁻⁵; 0.05-second input intervals. The controllability rank is unchanged across those perturbation sizes. See the measurement data and method.

The same-plant, model-derived LQR diagnostic holds both signed 5° double starts and signed 5°/15° triple starts without rail contact in the tested 60-second trials. Some larger double starts eventually hold after hitting the rails; those are not clean recovery results. A capable local reference helps separate a solvable holding problem from the difficulty of discovering its controller.

More acceleration did not fix a frozen policy

In the recorded g 10 compact-ReLU comparison, doubling acceleration broke previously successful small-angle holds even with the input units preserved. Widening the rail also changed the default position normalization: preserving the old encoding restored the small-angle trajectories. The 140° cases still failed. These are concrete calibration and strategy effects; they do not establish an energy shortage or an impossible plant.

The same task, a more informative start

A random-start NEAT run reached 1,910 generations without finding a usable local triple holder in this bounded experiment. A neural student fitted to a same-plant LQR teacher provided a stable starting policy. Full NEAT then improved disturbance rejection through real weight and topology changes.

Gravity1000, direct acceleration40000, original ±260px rail, linked 1–4° releases, fresh randomized tip impulses every3–5 seconds. Each passing trial keeps all three links above horizontal throughout with no rail contact. Physics120/240/480Hz; controller120Hz. No LQR teacher is called at neural playback. Impulse20 remains outside the refined policy’s reliable envelope.

This demonstrates a useful policy class and a successful initialization strategy for local holding. It does not prove that curriculum alone would find it from scratch, or that the learned holder can swing up. The raw-search and teacher-guided runs used different schedules and physical settings, so their training times are not a controlled comparison of initialization alone.

Gradients are another route to a controller

PPO does not need derivatives through the simulator. It learns from its trajectories and differentiates the network objectives. Replay uses the deterministic actor; the critic and random exploration are training tools. The implementation checks analytical gradients against finite differences and exposes the measured update signals. PPO · original paper

Would friction make swing-up easier?

Damping can help dissipate unwanted motion near a catch, but also removes energy during acquisition. The useful amount depends on the strategy. In this simulator, the damping control attenuates particle velocities; cart drag is a separate term. A fair ablation preserves damping per second when changing timestep.

Zero gravity is not a straightforward easier version of upright control. In our zero-command check, a static25° pose simply remained there: it already passed the above-horizontal criterion. A static140° pose remained below it. Neither result demonstrates an acquired balance strategy. Always compare against zero control before interpreting curriculum progress.

Compare online demos by their actual task

Force-driven examples do exist: DeepMind’s public cart-pole model uses a bounded force motor, and its two- and three-pole tasks start in swing-up configurations. Its ten-second default, observation, reward and plant differ from these90-second per-link tests. This portfolio’s evidence is the explicit interface, visible internals and reproducible limits; velocity control is not a universal explanation for other demonstrations. Actuator model · Task, initialization and reward

Plan the motion; learn to reproduce it

For a known plant, nonlinear trajectory optimization with feedback is a strong next comparator. Direct collocation, constraint-aware iLQR/DDP or constrained MPC can search coordinated motions explicitly; rail and actuator limits must still be checked on actual rollouts. A time-varying feedback controller can stabilize a feasible trajectory; a neural student can then learn from it.

This batch tested that route on the requested140°/98° linked start. A memoryless tanh student reached action MSE0.000915 on teacher labels but passed0/6 autonomous catches. Teacher-initialized PPO then used4.92 million physical steps: above-horizontal occupancy rose to72.4%, yet0/6 development trials held for the final30 seconds. These are failed development tests, not a hidden swing-up achievement. The sustained NEAT investigation screened117 high-gravity and54 teacher-initialized final candidates without finding a140° catch. The selected candidate from each branch then passed0/40 fresh90-second target trials.

Next, test trajectory phase or short memory against the same observation-only student, and collect corrections on the states the student actually visits. First make the same-plant teacher reliable under numerical refinement. A better curriculum is useful only if neighboring stages share a recoverable region; merely slowing the angle ramp does not establish that bridge.

From a local image to a mechanical decision

The hard-terrain controller must jump across a pit yet stay low under a spike ceiling. A changing course makes spatial information matter: the same frozen network faces new tile arrangements each trial.

All 416 parameters were learned with separable CMA-ES. This is a convolutional architecture trained without backpropagation. The inputs are simulator-provided occupancy, not camera pixels or a view of the whole course. The Perception circuit displays both sensor planes, every filter and pooled map, body channels, dense activations and output requests from the actual forward pass.

The eight failures comprise three falls, one hazard collision and four timeouts. The result establishes a working controller within this generator; it does not rank CNNs against every alternative. This platformer uses acceleration-limited drive, gravity and tile collisions. A jump request becomes an upward impulse only while grounded and outside cooldown. Network requests and the plant’s 30 Hz held commands are shown separately.

A circuit can be heterogeneous without being convolutional

The restored spatial-dodge study flattens two 16 × 16 danger maps, four body scalars and eight nearest-bullet tuples into 548 input neurons. Three hidden neurons and 2,744 evolved links connect them to two acceleration outputs. Those screen-shaped grids are real neural inputs; no convolution runs in this graph.

On 40 fresh moderate-density 20 s trials, 4 survived. Continuous tests registered 0.120 hits/s, versus 0.358 untrained and 0.719 stationary. Hiding images changed the result to 0.140 hits/s; hiding nearest-bullet state raised it to 0.509. The explicit motion channels contribute much more strongly here. It remains a visible proof of concept, separate from the stronger model-assisted dodge scorer and outside the automatic demo rotation.

Every heatmap and bar comes from the live controller. Warm and blue indicate signed magnitudes; maps scale per plane and bars clip at one. The live panels repaint alongside the main viewport. Optional Smoothing gently eases display intensity and bar length; exact numbers, choices, gates and timestamped history stay immediate. The motion-aware scorer still makes decisions at 20 Hz and records history at 10 Hz. Glowing glass and drifting embers are decorative. Collapse the circuit to stop its sampling and drawing, or choose Concise spacing for a shorter network view. Hover near a topology node to bring its label above the wiring.