Long-horizon robot tasks require retaining and using past information. We integrate Test-Time Training (TTT), which stores context in fast weights, into InternVLA-A1 and Pi0.5 to study action generation conditioned on human demonstrations through training on paired demonstrations and robot trajectories. We implement memory transfer between the two phases and truncated backpropagation, and examine training objectives, gradient coverage, and contrastive regularization across video, short-video, and single-frame cue tasks. In our experiments, adding TTT and training on paired data at the current scale did not lead to consistent target selection based on demonstrations. Under the tested settings, shorter sequences, full trajectory gradient connectivity, and contrastive regularization did not clearly improve responsiveness to demonstrations. When demonstrations were reduced to single-frame cues, real-robot execution still showed a preference for cups at fixed positions. We document the implementation, training configurations, and observations to support further inspection and replication.
Figures & tables
Figure 1: TTT layer architecture and memory operations. Green modules contain slow weights, while red modules represent a two-layer GELU MLP parameterized by fast weights. Normalization and projections transform the input into Q , K , and V . The right path predicts V from K and updates fast weights using the inner loss. The left path queries the updated model with Q , then applies an output projection and a gated residual connection. The two MLPs depict two calls to the same inner model, before and after the update. The batch dimension, head splitting, and inner normalization and residual details are omitted for clarity.
Figure 2: TBPTT and gradient accumulation in our implementation. Yellow boxes at the top are sampled decision anchors, each predicting an action chunk and computing its loss. Vertical dashed lines separate segments; yellow arrows indicate gradient contributions within segments and accumulation across segments. After the full trajectory batch is processed, accumulated gradients update slow weights, as indicated by the green arrow at the bottom. Inner fast-weight updates and state transfer across segments are not drawn separately.
Figure 3: Memory operations in human demonstrations and robot trajectories. Blue denotes the human phase, and green denotes the robot phase. Green arrows indicate fast-weight updates; yellow arrows indicate memory use in action-chunk prediction. The human phase may write more frequently, while the robot phase advances at action-chunk intervals. Fast weights carry across the phase boundary. Arrow density only illustrates that the frequencies can differ; it does not specify a fixed ratio.
Figure 4: Misalignment under two state-input conditions. In the upper GT-state case, earlier memory carries forward, but subsequent states still come from demonstrations rather than the execution of the model’s decisions. In the lower AR-state case, predicted actions feed into subsequent states, while visual observations still come from demonstrations and may conflict with the propagated states. The predicted trajectories and memory illustrate these relationships, not actual environment interactions during training.
Model
GT-state MAE ↓
AR-state MAE ↓
Real-robot completion rate ↑
baseline delta
0.023604
0.099308
90%
TTT delta
0.022631
0.091122
73%
TTT abs
0.028153
0.082412
64%
Table 1: Warm-up comparison after 6,000 training steps.
Data setting
Dataset size
InternVLA-A1 backbone pretraining
Approximately 692 million frames
Original video demonstration task
654 pairs, 556,302 frames; 621 training pairs
Short-video demonstration task
273 pairs; 264 training pairs
Single-frame cue task
210 pairs; 200 training pairs
Table 2: Backbone pretraining and demonstration-conditioned task data.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Incremental and absolute action targets during state autoregression. The gray straight line denotes the demonstration trajectory, and the dark gray curve denotes deviation caused by the previous action chunk. On the left, green arrows apply the original increments to the deviated state, potentially preserving the existing position error. On the right, red arrows keep absolute targets at the demonstrated positions. The figure illustrates target geometry, not learned corrective behavior during real execution.
Action representation
Training target relative to the ground-truth state
Absolute action if the same value is output after deviation
delta
0.0
0.7+0.0=0.7
abs
1.0
1.0
Appendix
Table 3: Action representations after state deviation.
Figure 6: Computation of the contrastive regularizers. First, each human demonstration is read, its terminal fast weights and graph are retained, and the memory separation regularizer is computed. Second, correct robot pairings are processed to compute the base training loss. Cross-pairings are then evaluated for the trajectory contrastive regularizer. Horizontal arrows indicate computation order; each robot branch restarts from the corresponding demonstration’s terminal memory rather than inheriting updates from the preceding robot branch.
Setting
Value
Learning rate
5×10−5
Learning-rate warm-up steps
50
Weight decay
0.01
Parallel trajectories per GPU
2
Generation loss weight
0.01
Fast-weight computation precision
FP32
Appendix
Table 4: Settings for the contrastive regularization comparison.
Training setup
Training GT-state MAE ↓
Training AR-state MAE ↓
Validation GT-state MAE ↓
Validation AR-state MAE ↓
Standard training (10,000 steps)
0.007650
0.024754
0.041284
0.107719
Contrastive regularization (10,000 steps)
0.008720
0.039244
0.044833
0.118272
Appendix
Table 5: Open-loop MAE after 10,000 training steps.
Training setup
GT-state
AR-state
With demonstration
Without demonstration
Demonstration benefit
With demonstration
Without demonstration
Demonstration benefit
Standard training (10,000 steps)
0.041284
0.041287
+0.0064%
0.107719
0.107573
− 0.1362%
Contrastive regularization (10,000 steps)
0.044833
0.044785
− 0.1072%
0.118272
0.118382
+0.0930%
Appendix
Table 6: Validation MAE with and without demonstrations after 10,000 training steps.
Training setup
Cup-selection success rate ↑
Standard training (10,000 steps)
25%
Contrastive regularization (10,000 steps)
12.5%
Appendix
Table 7: Real-robot cup-selection success after 10,000 training steps.
Post-training vision-language-action (VLA) models for specific robots and tasks requires in-domain demonstrations, yet collecting diverse robot data is costly. Egocentric human demonstrations provide a scalable alternative, but directly mixing human and robot data can introduce cross-embodiment discrepancies and degrade policy performance. To address this challenge, we introduce ReWeight, a framework that incorporates human data into VLA post-training through demonstration-level retrieval and sample-level weighting. ReWeight learns a cross-embodiment visuomotor representation that combines visual observations with future actions to measure behavioral similarity between human and robot demonstrations. Based on optimal transport, it retrieves human demonstrations relevant to the target robot data and assigns larger weights to samples with smaller cross-embodiment discrepancies. We evaluate ReWeight using π0.5 across eight simulation tasks and four real-world tasks under both clean and randomized settings. In simulation, ReWeight improves the average success rate of post-trained π0.5 from 39% with only robot data and 44% with randomly mixed human-robot data to 57%. In the physical experimental setting, it achieves an average success rate of 68.8%, outperforming the baselines by 28.8% and 13.8%, respectively. Overall, ReWeight provides an effective paradigm for transforming abundant egocentric human experience into transferable supervision for robot learning. (Project webpage: https://reweight-vla.github.io/)
Chenwei Wang, Dianye Huang, Match W. L. Ko +2
Department of mechanical Engineering, The University of Hong Kong, Hong Kong SAR, China · Institute of Artificial Intelligence (TeleAI), China Telecom, Shanghai 200232, China · Shenzhen Research Institute of Northwestern Polytechnical University, Xi’an 710072, China
Long-horizon robot manipulation requires a robot to both execute individual skills reliably and sequence them coherently over extended tasks. Most hierarchical vision-language-action (VLA) models make each such decision with a single forward pass, leaving no mechanism to allocate additional computation to difficult or consequential choices. We introduce τ0-VLA, a hierarchical robot foundation model that formulates high-level subtask generation as a compute-scalable inference problem through world-model-guided test-time computation. At each inference step, the high-level policy uses execution memory to generate a subtask and, when needed, searches over alternatives before committing to its output. A low-level policy then executes the generated subtask across multiple robot embodiments. The policy is trained on 40,115 hours of heterogeneous real-world data with multimodal co-training. Across in-domain and distribution-shifted settings, allocating additional test-time computation substantially improves next-subtask prediction accuracy, and these gains translate into higher closed-loop success on long-horizon robot manipulation tasks.
Xiaowei Cai, Yunuo Cai, Bingao Chen +36
Agibot Finch · Shanghai Innovation Institute · The Chinese University of Hong Kong
Vision-Language-Action (VLA) models trained on large-scale data have made remarkable progress, but they remain vulnerable to distribution shifts at deployment time. Recent VLA models suggest that prompts can serve as an efficient interface for steering policy behavior, but existing prompt-based steering typically relies on external guidance. This raises a natural question: can test-time training (TTT) for VLA be achieved by optimizing a prompt, so that the steering interface itself can be learned and adapted from interaction? We address this question with TTT-VLA, a test-time training framework based on Latent Prompt Optimization (LPO). During training, the latent prompt is learned with an additional proxy task, providing an extra learned conditioning signal for policy learning. At test time, TTT is performed by collecting interaction data from the current environment and optimizing only the latent prompt on those data using the proxy task's self-supervised signal, without modifying the policy itself. Experiments on SimplerEnv demonstrate that the proposed method consistently improves task success rates in both single- and multi-embodiment settings. Further analysis shows that the gains arise primarily from correcting a small number of critical decisions rather than globally altering policy behavior. These results suggest that LPO provides an effective and practical pathway for deployment-time improvement of foundation manipulation policies.
Wenbo Zhang, Jianxiong Li, Shuai Yang +4
ByteDance Seed · The University of Adelaide · Tsinghua University +3