7arXiv cs.LG (Machine Learning)·
RoboTTT scales robot policy context to 8K timesteps via Test-Time Training, enabling one-shot imitation and long-horizon tasks
NVIDIA researchers introduce RoboTTT, a robot foundation model training recipe that extends visuomotor context to 8,000 timesteps — three orders of magnitude beyond current state-of-the-art — without increasing inference latency. The approach integrates Test-Time Training into Vision-Language-Action policies, using fast weights (parameters updated by gradient descent during inference) to compress long histories into weight space. On real-robot manipulation tasks, RoboTTT achieves 87% performance improvement over single-step baselines and is the first system to fully complete a five-minute, ten-stage assembly task. The work identifies context length as a new scaling axis for robot foundation models, with 8K-context pretraining outperforming 1K-context by 62%.
Related guides (4)
Related events (8)
E-TTS: Embodied test-time scaling framework for robotic manipulation with history-aware iterative refinement
Researchers introduce E-TTS, a modular test-time scaling framework for robotic manipulation that unifies reasoning and action scaling via history-aware iterative refinement and vision-language verifiers. The framework addresses two gaps in prior work: underexplored reasoning scaling mechanisms and inadequate use of historical context in long-horizon sequential tasks. Evaluated across 4 benchmarks, 6 environments, 3 embodiments, and 4 base vision-language-action models, E-TTS achieves up to 33.14% improvement in simulation and 26.62% in real-world scenarios without additional expert data or retraining.
Test-Time Training End-to-End (TTT-E2E) Retrains Model Weights to Handle Long Inputs
Researchers from Astera Institute, Nvidia, Stanford, UC Berkeley, and UC San Diego introduced TTT-E2E, a method that compresses long context into transformer weights by training the model during inference via meta-learning. The approach uses sliding-window attention restricted to 8,000 tokens and updates only the fully connected layers of the last quarter of the network on each 1,000-token chunk at inference time, keeping per-token generation latency roughly constant as context scales to 128,000 tokens. TTT-E2E slightly outperforms vanilla transformers on next-token prediction loss across long contexts and matches efficient architectures like Mamba 2 and Gated DeltaNet on inference speed, but fails dramatically on Needle-in-a-Haystack retrieval beyond 8,000 tokens and incurs substantially higher training latency. The work reframes long-context handling as a training-inference trade-off rather than an architectural design problem.
TREAD: VLM-based re-labelling framework improves robot policy generalization via dataset augmentation
TREAD (Task Robustness via Re-Labelling Vision-Action Robot Data) is a scalable framework that uses pretrained Vision-Language Models to augment existing robotics datasets without new data collection. The approach decomposes demonstrations into sub-tasks, segments videos accordingly, and generates linguistically diverse instruction labels, enriching language-action pair diversity. Evaluations on the LIBERO benchmark show improved generalization to novel tasks and goals, addressing a key limitation of current robot learning policies.
RoboReward: Vision-Language Reward Models for Robot Training via RL
Researchers at Stanford and UC Berkeley developed RoboReward, a family of 4B and 8B vision-language reward models designed to provide reward signals for robot reinforcement learning across diverse robot types and tasks. The team built a novel dataset by augmenting successful robot demonstrations with synthetically generated failure examples using GPT-5 mini and Qwen3-4B, then fine-tuned Qwen3-VL models to predict task progress scores. RoboReward 8B outperformed GPT-5, GPT-5 mini, and Gemini Robotics-ER 1.5 on the new RoboRewardBench evaluation, and in real-world robot trials substantially exceeded prior reward model baselines while still falling short of human-assigned rewards. The authors also release RoboRewardBench as a community benchmark for reward model evaluation.
Self-Guided Test-Time Training improves long-context LLM utilization by up to 15%
Researchers propose Self-Guided Test-Time Training (S-TTT), a method that addresses the degradation in accuracy LLMs exhibit on long inputs by having the model first identify relevant evidence spans before applying test-time training only to those spans. A preliminary study on LongBench-v2 shows that TTT on randomly sampled spans hurts performance while TTT on oracle spans substantially helps, motivating the self-guided selection approach. Evaluated on LongBench-v2 and LongBench-Pro with Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, S-TTT achieves up to 15% relative accuracy improvement. The method offers a practical path to better long-context utilization without the prohibitive cost of full-context adaptation.
FORCE: Efficient RL fine-tuning for Vision-Language-Action models via value-calibrated warm-up and self-distillation
Researchers introduce FORCE, a 3-stage reinforcement learning fine-tuning framework for Vision-Language-Action (VLA) models that addresses sample inefficiency caused by unstable Q-functions and low-quality exploration data. The framework uses a Value-Calibrated Warm-Up phase followed by Q-function-filtered policy updates, eliminating the need for costly human interventions during training. Evaluated on simulation and real-world robotic tasks, FORCE achieves a 79% absolute improvement in task success rates, outperforms prior RL methods by 10%, and accelerates training by 32.5%.
VLK: Synthetic vision-language-kinematics supervision enables humanoid loco-manipulation from egocentric observations
Researchers introduce a pipeline that generates 48,000 paired vision-language-kinematics trajectories synthetically using 3D Gaussian Splatting to reconstruct indoor scenes, bypassing the need for expensive human-annotated robot data. A VLK policy trained on this data predicts whole-body kinematic trajectories from egocentric images and language instructions, which a whole-body tracker converts to physical actions. The approach is validated on a Unitree G1 humanoid performing navigation and object transport, demonstrating viable sim-to-real transfer for perception-based loco-manipulation.
TempoVLA: Speed-Controllable Vision-Language-Action Policy for Robot Manipulation
Researchers introduce TempoVLA, a Vision-Language-Action model that enables explicit speed control during robot manipulation by conditioning on a speed signal rather than inheriting a fixed speed from training data. The system pairs Variable-Speed Trajectory Augmentation (VSTA), which re-times demonstrations by merging or splitting actions, with a model-side conditioning mechanism. Experiments in simulation and real-world tasks show flexible bidirectional speed control, with dynamic adaptation—accelerating in low-risk transit phases and decelerating for high-risk contact stages—achieved by coupling with a large multimodal model.