MetaHuman update: trained 2B specialists and clearer task outcomes

The latest MetaHuman work changes both the models and the software around them. Trained Qwen3.5 2B Intent and Task adapters replace the earlier 0.8B production specialists. An independently trained 9B autonomy adapter has an explicitly experimental assignment. Monitor and task-lifecycle repairs address unnecessary server work and failed actions that could leave misleading unfinished goals.

These are separate advances. A faster response, a better training score and a correct physical action each need their own evidence.

The small models are still trained specialists#

Intent interprets a request and selects its route; Task selects supported actions and their controls. Both now use separate trained adapters on one shared 2B base. The architecture retains specialization rather than sending every task to the larger conversational model. The 9B model continues to handle conversation and vision; model roles remain configurable in MetaHuman.

The October 15 comparison used identical evaluation requests:

Automated reference agreement	Previous 0.8B deployment	2B candidate
Intent	219 / 297	266 / 297
Task	578 / 1,293	1,194 / 1,293

These counts measure agreement with reference decisions, not semantic or robot success rates. Review still found at least 21 previously valid Task decisions that regressed, including issues with conditions, counts, order and movement parameters. The owner selected the new adapters with these limitations recorded.

Production uses a Q8_0 2B base with the trained F16 adapters. Measured GPU allocation fell from 5,684 to 3,982 MiB compared with the BF16 configuration. Forty paired precision checks preserved the selected programs and task controls; 38 parsed outputs were identical and two differed in wording. This was a focused comparison, not a complete accuracy or physical-behavior benchmark.

A separate experiment in autonomy#

The earlier 9B diagnostic report recorded no production promotion. A later owner decision approved experimental use of the independently trained autonomy adapter through its own Autonomy category. Active Operator and the Boredom Reflection/Movement planners use that category. Conversation and the image-observer assignment remain unchanged.

The small serving comparison was mixed: the base produced valid outputs in eight of eight cases, the adapter in seven of eight. Median complete-call time fell from 37.7 to 3.1 seconds because the adapter produced far fewer tokens, not because it generated each token faster. Neither that comparison nor the synthetic-image smoke test establishes reliable autonomous behavior.

This adapter is distinct from the longer-term goal of a learned conversational persona. The current persona still uses contextual identity instructions.

Removing work that made the interface sluggish#

A live CPU profile found repeated Agent Monitor processing dominating the application thread: the monitor repeatedly read and parsed the entire daily audit log while building status views for agents.

The repair removed automatic historical-log processing and its retired UI fields from that owner. Live service, process, bridge and lifecycle information remain available, along with explicit manual diagnostics. Recorded source snapshots dropped to roughly 3 ms after warm-up, with zero audit reads.

The repaired application was activated without restarting the model or voice services. Whole HTTP requests still varied with workload; this result does not claim that every source of conversational latency has been fixed.

A rejected action should retain its actual result#

One recorded reflection selected an unrelated Stand command. The body rejected it, but the task layer converted that terminal failure into continued work and eventually retained a goal waiting for input.

The repaired task owner preserves a terminal failure for a one-off action instead of creating another goal-review cycle. A subsequent change removed the remaining task-level continue mechanism. Goal Review assesses evidence and updates status; it does not select more actions or speak. New work comes from human input or independently admitted autonomy. An already selected program can still advance through its steps.

Software acceptance checks verified that the rejected one-off case ends as failed, gives one reply and does not repeat the action on re-entry. A genuine waiting goal can coexist with newly admitted work. The application changes were activated October 16. These checks used simulated actions, so they do not establish physical task success or repair every model-selection error.

Seeing now, remembering later#

Tests with the body disassembled exercised the normal conversation route, without movement. A fresh camera image reached the 9B Persona model, and one reply matched the scene. Another case exposed stale status being described as recent. A camera-freshness repair was activated, but an unsupported identity guess in a later response shows that visual grounding still needs work.

The memory index was rebuilt to include both sides of conversations rather than only the user’s messages. That restored the assistant’s replies to the index. Retrieval could still rank a similar old question above its answer, and an ordinary recall request could be routed to a new camera capture. Storing the conversation and recalling the right evidence remain different capabilities.

Revised memory-routing training data and evaluation checks were prepared as dataset version 20. That preparation is not a trained or deployed new adapter. Reliable recall, correct visual interpretation and complete autonomous outcomes remain open qualification work.

Read the evidence alongside the features#

This update follows the Robot Bible and October 16 workflow investigation records inspected on October 18. The source snapshot records the local documents used, including unpublished changes. Private conversation content is not reproduced here. The companion guide explains the intended experience; current status keeps the latest qualification boundaries visible.