The model learns what might happen. Its current planner is mostly trying to copy an old body pose.
That is the main disconnect from your goal: a body that learns to get closer to something it wants, reach it, grasp it and keep hold of it.
Wellness is the computer's score for how well the body is doing. It includes touch comfort, ability to move and how accurately the model predicted its senses.
The parts named wellness, imagination, intention and forward model exist. But several connections needed to make them work together are missing, switched off or bypassed. We should have checked that complete chain before treating another small reaching repair as the route to the larger goal.
Update, 7:00 PM: The new code passed 14 checks with a small test model. It examined predicted moves without changing what the model had learned. Next: check MIMo's real joint setup and connect the scores for sensing, movement and improvement from practice. Claude is doing this. The current movement chooser is unchanged.
What you expected—and what the code currently does
INTENDED GENERAL LEARNING LOOP
CURRENT BODY-PLANNING ROUTE
The missing connection: the active body planner does not choose a sequence because of the useful consequences expected later. Choosing a good past moment is different from judging the future results of a proposed sequence.
Example only: shifting the body first might briefly move the hand away from a ball, but make the next reach possible. A rule that demands an immediately closer hand can reject that first move. The model should discover the useful sequence; we should not prescribe “shift, then reach, then close fingers.”
Where the connections fail—and one risk to holding
1. “Wellness drive” does not predict the result of a move
Only a fixed body-position penalty supplies this input to the controller. The measured touch, physical capacity and prediction-accuracy terms supply no direction for changing the action. The input can push away from the penalty's position boundary; it cannot answer “will this move improve comfort later?”
The term is named motor_state. It measures body-point positions relative to their centre, not actual joint-angle limits. wellness.py:310–393; lem_model.py:3343–3451, 4358. Its slope is an input, not a learning signal for future benefit. The separate future-wellness learning branch at 3459–3467 is off.
2. The body score leaves future wellness out
The active search tries to reduce the predicted difference from the remembered pose. Even enabling the existing future-wellness option would not repair this route: its physical-goal optimization bypasses that term.
intention_generator.py:270–287, 609–627. The comparison also requires improvement within the quarter-second window.
3. Later imagined moves use an old body position
The prediction loop advances its sensory estimates, but carries the starting physical pose forward unchanged. A second wrist-relative command can therefore be interpreted from where the wrist started, rather than where the first move put it. The current planner limits itself to one move to avoid relying on that error.
predictive_state.py:173–174, 246–253; lem_model.py:6993–7012.
4. Practicing predictions is not choosing a sequence
Imagination practices predicting three successive moments using movements that were actually performed. That improves a prediction skill. It does not, by itself, compare different possible sequences and choose the one with the best later result.
live_test.py:247–304; lem_model.py:3497–3581. Three prediction steps span 0.15 seconds of simulated time here.
5. The original plan also needs a correction
The original specification's “multi-step” example chooses the next move toward the same goal, then repeats. It never compares the total consequences of competing sequences. Restoring that example alone would leave this weakness.
UEMA specification 3.6, §10.3, lines 534–547. Its §7.4 also omits the connection from a hypothetical command to its future physical consequences.
6. Learned future-condition forecasts do not choose
The persistent-memory model learns to forecast bodily needs. The self-state model also learns future body condition. Their forecast outputs are not used by the active command chooser. Persistent memory still helps the forward model; the missing connection is specifically from those future-condition forecasts to the action decision.
lem_model.py:4301–4318, 4841–4851; self-choice off.
A holding risk in the goal score
The goal score emphasizes how each signal differs from its recent average. If a good condition stays steady, its difference from that average shrinks. That could make maintaining a useful condition stop standing out. The formula confirms this risk; it has not yet shown whether the body abandons a real grasp for that reason.
wellness_channels.py:380, 448–483; lem_model.py:9191. Repair review must distinguish “getting better” from “staying well.”
A separate problem with the ball example
The current caregiver ball can be made to follow the hand by code after contact. That is useful assisted experience, but it cannot tell us whether finger forces actually lifted and held the ball. Free-moving objects already exist in the environment; genuine grasping must be judged on physical support.
caregiver.py:428–465; active option: toy-hand-off follow. Keep assisted records labelled as assisted.
What this means for the repair
Keep the learned model and repair how it uses its predictions to choose. This does not call for a programmed list of toddler actions or a reward for “grasp the ball.”
- Make imagined body changes usable by the next imagined move. The second move must start from the first move's predicted result. Joint geometry can keep a pose consistent; the learned model still has to predict motion, contact and support.
- Connect predicted consequences to the existing wellness system. Candidate actions must change predicted future senses and bodily condition, and those changes must reach the score used to choose. A detached measurement of the current pose cannot supply that connection.
- Compare sequences, not just successive single moves. Judge a first move partly by what it enables afterward. Execute only the first part, observe again and revise the rest.
- Teach and judge the same actions that reach the body. Include joint limits, hand-relative coordinates, smoothing, command duration and reflex interventions. Do not compare an imagined command with a differently executed one.
Why would it want the ball? Recognizing or predicting a ball is not enough. The learner needs experience connecting interaction with things to its own needs, curiosity or control. Those learned connections must influence imagined choices. We should not insert a special “ball reward” to fill that gap.
Also required: keep useful conditions valuable after the initial improvement, and preserve opportunities to discover new actions. A score that forgets the value of continued holding, or a schedule that crowds out exploration, would undermine the repaired planner.
First implementation decision: reuse the existing prediction loop and sequence chooser where their interfaces fit. Review the exact replacement for stale physical state and the shared consequence score before enabling either. The current short pose-continuity package does not repair these connections.
First decisive check: demonstrate that a candidate's first action changes a later predicted consequence and that this later consequence can change which first action is chosen. Then compare predictions with real outcomes during preserved-state learning. A simple artificial example can check the computation; only physical behavior can establish the ability.
Source/design owner: Codex. Independent review and implementation packaging: Claude. No training or standalone diagnostic is authorized by this page. Existing cumulative diagnostic limit: 15 minutes per named question. No estimate for acquiring the whole ability is available from these source findings.
The whole model: what each part receives, learns and affects
This covers the active H84 pipeline and the related modules that exist but are inactive. It is a source and saved-configuration review, not a claim that every parameter has been experimentally validated.
| Part | What enters / what is learned | What it actually affects now |
|---|---|---|
| Senses and encoders | Camera images, skin contact, sound, body position, issued commands and balance signals. Each is compressed to a 192-number vector. | These vectors feed the forward model. Vision can carry object information; lack of a named object tracker does not prove the model cannot represent a ball. Compression quality still matters. |
| Forward model | Six input streams and a context of 128 model moments. Four attention blocks learn relations across senses and time. | Predicts future sensory vectors. An explicit command input and separate body/gaze heads help commands influence predictions. This is useful learned machinery. |
| Prediction heads | Model output, current sensory vectors and direct action paths. Delayed lessons compare forecasts with later real observations. | Supply predicted senses and physical body changes. Predicting separate body points does not automatically enforce connected bones or floor contact. |
| Imagination training | Recorded actions; predicted sensory state is fed into the next prediction. Three-step errors train the model. | Practices forecasting. It is active, but the body planner's one-command restriction prevents its earned longer horizon being used for that choice. |
| Measured wellness | Prediction accuracy, touch comfort, physical capacity, body coherence and motor-state formulas, plus separately ranked movement/agency/learning channels. | Values observed moments. Goal ranking subtracts each body's recent average from each scored channel, then normalizes and weights those differences. Pain/harm limits are checked separately. The motor gradient has the current-pose limitation described above. |
| Learned wellness estimator | Actual sensory vectors paired with measured wellness. | Learns to estimate wellness from compact sensory descriptions. Available for predicted states, but its future score does not select the active body command. |
| Curiosity and learning progress | Recent prediction error, reduction in error, and control evidence. | Change how past moments are ranked. Current curiosity does not search imagined sequences for the most informative later experience. |
| Recent goal memory | Last 200 model moments: sensory vectors, scores and associated body poses. | Selects a valued remembered moment. At 0.05 seconds per moment this covers 10 seconds. The completed run planned every 0.5 seconds; the pending package would plan every 0.25 seconds and hold its selected goal for 3 seconds. |
| Persistent memory | Longer lived-event history and recurrent carry; trained to predict bodily needs under issued commands. | Biases the forward model. Its separate needs forecast is not read by the chooser. This longer memory is distinct from the recent goal list; neither alone implements choosing for a delayed benefit. |
| Intention / body planner | A remembered goal, current state and candidate commands. | Optimizes the predicted pose after one 250 ms command. In the pending package it would keep a goal for 3 seconds while replanning; that package has not run. |
| Hand planner and FK/IK | Remembered hand shape and joint geometry. FK calculates where a chain of joints puts the fingers; IK searches joint bends for requested positions. | The hand stage fits a remembered finger shape without a contact/object objective. It makes finger commands compatible with hand joints; it does not predict whether contact forces will hold an object. Whole-body future dynamics still need their own valid prediction. |
| Motor controller | Model state, achieved pose and the current wellness gradient. In baseline mode, its output is trained toward the next observed pose. | Supplies movement proposals when the planner does not replace them. Extra network depth means more calculation, not a longer time horizon. |
| Command execution | Proposal/planner choice, joint limits, strength limits, smoothing, finger servos and held-command timing. | Produces the actual commands sent to physics. Exploration noise is skipped when a planner command exists. The pending schedule aligns planning with every command latch, so it could greatly reduce executed exploration. The completed run planned at every second latch. Actual command-source records decide how often each acted. |
| Environment and reflexes | Fixed joint servos, passive springs, fatigue-dependent force, touch precision, gaze reflexes, spontaneous eye movements and caregiver behavior. | These alter movement and sensation without being learned action choices. The ball-following assistance is one example. Their effects belong in the forecast and in the record of what caused a behavior. |
| Body schema and entity self | Correlations between the body's movements, body parts and entities; evidence-tested links to wellness. | Body schema contributes coherence. The entity-value reader can alter goal value only when its evidence gates pass. Some entity inputs come from simulator metadata and must not be mistaken for discoveries from camera pixels alone. |
| Self-state model / sequence chooser | Body condition and action sequences. Self-state prediction learns in this configuration. | The chooser is off. Its existing implementation also interprets later wrist-relative actions using the starting pose. Enabling it alone would preserve that error. |
| Contact predictions | An active head learns where contact will occur. A different, longer contact-consequence module is off. | The active contact head contributes control evidence to goal value. The longer module's sequence forecast refuses the current mechanical contact profile because required pressure-history outputs are missing. |
| Motor synergies / latent body hierarchy | Coordinated movement representations and recorded physical changes. | Synergies learn but are disabled for planner and exploration use. The separate body hierarchy runs in shadow: it observes/learns without choosing the body's moves. |
| Anticipation, speech and many minds | Some related modules or interfaces exist; not all are active skills. | Anticipatory gaze supervision is off. The legacy voice output has no voice target in this trainer. No active many-minds skill-splitting/arbitration system appears in the inspected H84 route. |
| Learning, replay and preservation | Both same-moment reconstruction and next-moment prediction lessons remain active, alongside physical-consequence, wellness-estimation and self-state lessons. Replay reuses recorded experience. | The reconstruction and future-prediction lessons share parameters; their relative influence needs checking before blaming only the planner. Weights update and complete checkpoints preserve model, optimizer and worlds. This preserves learning; it does not prove a skill has been retained or transferred. Those require ability evidence. |
When the working route narrowed
- Original specification: the planner reduces distance to a desired sensory state. Its multi-step example greedily repeats that operation. The stronger sequence-selection behavior was asserted, not supplied by that example.
- February 18:
5673aba25already chose an action by sensory-goal distance and strength cost; later steps were planned greedily. Wellness could stop a plan. - August 20:
4eecc50aewired the recent-memory goal selector into the live bridge. Wellness picked the remembered goal. - August 25:
873a36d4afrestricted fixed-horizon planning to one step because later commands reused the original achieved pose. This is not the clamp controlling today's earned-horizon configuration; the September restrictions below are. - September 2:
089bf13daadded an earned horizon: allow more imagined commands when prediction evidence supports them. This still runs, but the active physical/hand restrictions override its longer horizon. - September 9: prospective wellness was added behind an option, off by default. It remains off in the examined command. Claude's historical scan found it off in all 181 experiment argument records and 33 preserved argument records that contained the field. That is the inspected inventory, not proof about every run ever made.
- September 24:
ecfc9060fintroduced the optional physical-memory route and physical/hand one-command restrictions. Wrist-relative hands entered production then; the general physical-memory route was initially kept off. - October 22:
d4a3479a65made physical pose return the search priority and required immediate pose improvement. H84 ran that route under a Codex repair decision on October 22. It improved the definition of a narrow pose test while moving the active decision further from general future-consequence planning.
Earlier two- or three-step plans could veto a first move, but did not optimize that first move using the later result. Multi-step training could still improve the shared prediction weights. The documented prospective-wellness test was stopped by the operator on October 5; this audit does not restart it. A repair must connect and validate the full loop, not merely enable an old option.
Evidence, uncertainty and protection of existing learning
Examined source: D:/UEMA_wt/persistent_goal_20261025, commit 2d7718add67e7ad8a461a6f7cbbccf5cee69e6f6. Completed-run receipt: target_unblended_496250_u_a2, endpoint 496256. Pending command: persistent_goal_20261025/package_pg/command_pg.json. The pending goal-persistence change is not a completed training result.
Search size is not movement distance: the ordinary optimizer can change each eligible position coordinate by up to 1.80 cm and rotation by up to 11° relative to its first proposal. Those bounds do not limit how far the body can physically move. The enabled stored-target candidate is outside that search box. It still must beat the first proposal's forecast and the actual starting pose. “First proposal” does not mean “do nothing.” See physical_goal.py:54–59, 291–298, 325–363.
Two shared learning objectives: lem_model.py:3327–3332 constructs a same-step reconstruction target; 3611–3619 constructs the delayed next-step target; 3871–3899 includes both in the loss. Their coexistence is verified. Whether they compete enough to impair useful forecasts remains a measurement question. Read existing loss/gradient records before changing their weights.
The completed receipt lists about 18.1 million parameters and the six sensory/action streams above. Its recorded command at step 496250 came from the planner. This audit inspected records and source without loading or modifying the learned weights.
What remains usable: observed transitions, sensory prediction training, physical-command lessons and preserved checkpoints. Narrow pose improvements remain narrow pose results. Assisted ball-following records remain assisted experience. No percentage of “lost training” can be inferred from these code findings.
What needs further learning: a repaired future-state interface and any newly connected consequence forecast must be trained and checked on actual outcomes. Existing experience should be reused where its inputs and targets still match. Preserve the eight worlds and full optimizer/model state; do not restart the learner to simplify the comparison.
What this review does not claim: one-step replanning can produce longer behavior when each local move helps. Exploration can discover new poses. Visual vectors can represent objects without named slots. None of those facts supplies the missing general comparison of delayed consequences. Likewise, a free root joint permits translation but does not demonstrate successful crawling under these actuators and contacts.
Evidence that will settle the repair: the same candidate action must affect later forecasts, later wellness must affect the chosen first action, and the observed physical outcomes must improve under the resulting choices. Keep the external ability measurement separate from the model's own training objective.
- Original model specification — §§7.4, 10.3, 11, 13–18.
- Gain-of-function v4 — already distinguishes prediction from useful control and warns against assuming a long horizon.
- Current work record — model_connection_audit; source, review, follow-up and file disposition.
- Independent source challenge and reconciliation.
Adversarial review: what must not slip through again
- Trace a complete path from a candidate action to its later physical consequences, to wellness, to the chosen command. A module name is not evidence that the connection exists.
- Check the exact live configuration and the executed command. Distinguish trained, used for choice, shadow-only and switched-off modules.
- Advance the entire state needed by the next move, including body geometry, command duration and relevant memory. Never reuse the starting pose as a future pose.
- Include a sequence whose useful first move temporarily worsens the immediate pose score. Explain how the selector can choose it.
- Keep candidate forecasts comparable with the actual body: units, coordinate frames, joint constraints, smoothing and interventions must agree.
- Challenge the wellness objective itself: can stillness, repetitive touch or another useless loop score well? A better predictor alone cannot correct a poor objective.
- Judge useful, repeatable behavior with real contact physics. Do not substitute a lower prediction error, a drawn pose or caregiver assistance.
- Preserve learned state and identify which old observations remain valid. Make runtime and temporary-file costs explicit before execution.
Independent source review received from Claude at 16:19 PDT, with a history addendum reconciled afterward. Corrections added the unused future-condition forecasts, goal-score averaging, distinct hand/body/contact scores, exploration scheduling and environment effects. Reviewer overstatements rejected: a 1.8 cm search adjustment is not an actual movement limit; the first proposal is not “do nothing”; pending schedule is not completed-run behavior. Draft self-review also corrected three overclaims: an explicit object tracker is not mandatory; one-step planning is not proof that all longer behavior is impossible; the original specification also lacks joint sequence optimization. A second review caught damaged punctuation introduced while editing; those passages were rewritten directly as UTF-8. “Moment” means an ordinary recorded instant here; the technical time unit is a model step, 0.05 seconds in this run.