HumanoidTTT: Test-Time Capability Reuse for Efficient Humanoid Control

HumanoidTTT introduces a framework for test-time capability reuse in continual humanoid control, addressing the challenges of reliable motion reuse under changing robot states and managing validated capabilities within finite storage. The core contribution is enabling direct reuse of qualified motions only when the current state satisfies entry certificates, while adaptively retaining useful capabilities through deployment feedback.

The gist: HumanoidTTT enables reliable and efficient reuse of validated complete motions while adaptively retaining useful capabilities through online consolidation.

Selective Full-Motion Reuse

This mechanism authorizes the direct reuse of validated complete motions only from certified applicable entry states, allowing accepted reuse to bypass fresh generation. Each stored capability consists of a validated complete motion and an applicability certificate defining its Entry Applicability set, which is the set of admissible robot entry states from which the motion can be reliably reused. For each recurring request, a matched capability is directly reused only when the current entry state falls within this set, allowing the stored complete motion to bypass the Frozen Motion Generator and proceed directly to execution; otherwise, it falls back to fresh motion generation followed by execution-based qualification.

Test-Time Capability Consolidation

This process adapts which qualified capabilities persist in a bounded Full-Motion Store using subsequent deployment reuse as feedback. When newly qualified capabilities compete for limited storage, the consolidation policy decides whether to skip the new capability or replace an existing one, and adapts its retention decisions using subsequent reuse utility observed during deployment. The policy is invoked exactly when a qualified miss creates competition for a full Store, expressed by the condition where a qualified request arrives at a full Store. The online network selects either SKIP or REPLACE(j) from the current Store-conditioned state, and the reward is defined as execution-successful Store reuses/observed requests.

System Components and Logic

The framework separates continual capability reuse into two design principles: a read-side applicability problem and a write-side retention problem. The read path involves checking for an exact capability/signature match against the stored capabilities, requiring fixed physical checks and candidate-relative certificate membership, defined by the formula: Accept(i) = Match(qt, di) ∧ Phys(st, Ht) ∧ x(i)t ∈ Ai. If a match is found and the physical criteria are met, an accepted motion is retrieved for direct execution without motion generation.

Kernel-Aware Real-Time Deployment

HumanoidTTT utilizes a heterogeneous CPU–GPU execution path to separate frequent lightweight capability management from compute-intensive motion generation. For accepted reuse requests, Full-Motion Store lookup, Entry Applicability evaluation, complete-motion retrieval, and Store writeback remain on the CPU without activating the motion generator. Capacity-constrained test-time consolidation also runs in CPU FP32, where the compact DoubleDQN is invoked only when a qualified capability competes for a full Store. This event-driven execution keeps online adaptation off the critical GPU generation path and avoids making test-time learning a fixed cost for every request.

Experimental Results

Experiments demonstrate zero unsafe accepts and a 16.4× end-to-end speedup over fresh generation, while online consolidation improves avoided generator calls by 13.2 per 200 requests over its frozen counterpart. HumanoidTTT reduces median preparation latency from 5057.996 ms to 29.514 ms on accepted reuse hits, corresponding to a 171.4× speedup, and achieves a 16.4× end-to-end speedup over fresh generation. Under a capacity-10 Full-Motion Store, online consolidation avoids 13.2 additional generator calls per 200-request stream compared with the same pretrained retention policy with online updates disabled.

Contributions

The main contributions are:

1. Introducing Selective Full-Motion Reuse, which turns execution-qualified complete motions into persistent capabilities and enables reliable direct reuse across admissible entry states, bypassing fresh motion generation on accepted reuse hits.

2. Introducing Test-Time Capability Consolidation, which learns which qualified capabilities should persist in a bounded Full-Motion Store based on subsequent reuse utility observed during deployment, while keeping motion generation and acceptance criteria fixed and freezing each applicability certificate after acquisition.

3. Systematically evaluating HumanoidTTT in terms of full-motion reuse reliability, accepted-hit efficiency, end-to-end inference efficiency, and finite-capacity capability retention.

Limitations

HumanoidTTT currently freezes each Entry Applicability certificate after capability acquisition, leaving persistent changes in robot dynamics, environmental conditions, or admissible entry regions outside the current adaptation scope. Future work will extend Entry Applicability to update from deployment experience and evaluate the resulting capability reuse and finite-capacity consolidation over longer and more diverse real-world deployments.

Here are specific improvements for existing AI systems based on the HumanoidTTT framework, detailing what these improved systems can achieve:

1. The ability to execute previously validated full-motion sequences reliably across changing robot states without requiring a complete motion regeneration pipeline.

2. The capability to serve recurring motion requests with a significant reduction in end-to-end latency (up to 16.4× faster than fresh generation) by bypassing the Motion Generator entirely on accepted reuse hits, leading to near-instantaneous execution of known, safe motions (median preparation latency drop from 5058ms to 29.5ms).

3. The implementation of a dynamic, data-driven memory management system (Test-Time Capability Consolidation) that learns which previously executed motions are most useful for future requests and adaptively retains them in a bounded store, optimizing storage utilization while maximizing the potential for fast reuse.

4. The creation of a robust safety mechanism that guarantees zero unsafe accepts during reuse, ensuring that only motions whose execution is certified to be safe from the current robot entry state are played back, effectively eliminating the risk of blindly replaying potentially unsafe previous behaviors.

5. The enhancement of real-time inference efficiency by employing a heterogeneous CPU–GPU execution path: lightweight capability management and safety checks run on the CPU (FP32), while only compute-intensive motion generation is dispatched to the GPU (via ONNX Runtime CUDA) when a fresh motion is required, drastically reducing critical-path computation for recurrent requests.

6. The development of an adaptive, frozen certificate system where the entry applicability set for each stored motion is fixed after acquisition but can be leveraged by subsequent deployment experience to improve future retention decisions via online reinforcement learning (Double-DQN), allowing the system to learn which capabilities are worth keeping without risking the correctness of the motion generation or acceptance criteria themselves.

Recent advances in motion generation and whole-body tracking have enabled humanoid robots to execute increasingly diverse motions, yet the same motion capabilities may be requested repeatedly during continual deployment. Reliable reuse is challenging because intervening motions can change the robot's entry state, making previously successful motions unsafe to replay blindly. Meanwhile, validated capabilities accumulate during deployment, while bounded storage requires deciding which ones are worth retaining. To address these challenges, we present HumanoidTTT, a framework for test-time capability reuse in continual humanoid control. Specifically, we introduce Selective Full-Motion Reuse, which authorizes direct reuse of validated complete motions only from certified applicable entry states, allowing accepted reuse to bypass fresh generation. We further introduce Test-Time Capability Consolidation, which adapts which qualified capabilities persist in a bounded Full-Motion Store using subsequent deployment reuse as feedback. Experiments demonstrate zero unsafe accepts and a 16.4 times end-to-end speedup over fresh generation, while online consolidation improves avoided generator calls by 13.2 per 200 requests over its frozen counterpart. Overall, HumanoidTTT enables reliable and efficient reuse of validated motion capabilities while adaptively retaining useful capabilities throughout continual deployment. Code: https://github.com/AIGeeksGroup/HumanoidTTT. Website: https://aigeeksgroup.github.io/HumanoidTTT.