6arXiv cs.LG (Machine Learning)·
πR² enables reactive real-time replanning for robot manipulation flow policies
Researchers introduce πR², a method that makes large action-chunking flow policies reactive and real-time by splitting conditioning into fast proprioception and slow vision-language channels, plus a latency-adaptive denoising schedule. Applied to NVIDIA's GR00T-N1.7 on a real xArm6+XHand platform, it achieves ~25Hz closed-loop replanning (roughly 4× faster than the base policy), acting on fresh observations every 40ms. The approach improves task success rates by up to 23% in simulation and 30% in real-world manipulation over the strongest baseline, requiring only minimal architectural modification and fine-tuning from a pretrained checkpoint.
Related guides (4)
Related events (8)
PACT: Hybrid SLM deliberation architecture improves reactive RL policies in unfamiliar environments
Researchers propose PACT (Plan, Align, Commit, Think), a hybrid architecture pairing a fast reactive RL policy with an asynchronous small language model planner for deliberation. The SLM generates and validates candidate action plans via simulation before committing to execution, bypassing the RL policy without retraining. Evaluated on FrozenLake configurations of increasing difficulty, PACT outperforms baselines using only a 2B-parameter SLM, suggesting complementary strengths between deliberative planning and reactive execution.
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%.
REGRIND: Minimalist retargeting-guided RL pipeline for dexterous manipulation from single human demo
Researchers introduce REGRIND, a pipeline that learns dexterous robot manipulation policies from a single human demonstration by retargeting hand-object motion to robot kinematic references and training a residual RL policy to track object-centric keypoints. The system achieves zero-shot sim-to-real transfer on multi-fingered hands for contact-rich tasks including scissors operation and screwdriver turning. The paper systematically analyzes key factors governing sim-to-real transfer in dexterous manipulation, extending a recipe previously validated for humanoid whole-body control into the harder contact-rich manipulation domain.
π0 and π0-FAST: Vision-Language-Action Models for General Robot Control
Hugging Face published a blog post covering π0 and π0-FAST, vision-language-action (VLA) models developed for general-purpose robot control. These models combine vision and language understanding with action generation to enable robots to perform a broad range of manipulation tasks. The post appears to be a technical overview or release commentary on Physical Intelligence's robotics foundation models, situating them within the broader VLA research landscape.
ACPO: Adaptive Clip Policy Optimization improves RLVR training for LLM reasoning
A new arXiv preprint provides theoretical analysis of Reinforcement Learning from Verifiable Rewards (RLVR) updates, identifying off-policy degree and gradient expectation as key factors governing update dynamics. The authors show that differences in gradient steps per rollout substantially affect importance sampling ratio distributions and which tokens dominate updates. Based on this analysis, they propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries per token group by empirical variance of importance sampling ratios, outperforming DAPO and CISPO baselines on 3B and 7B models across math, tabular QA, and logic benchmarks.
PAC-ACT: Reinforcement learning post-training framework for Action Chunking Transformer robot policies
PAC-ACT is a reinforcement learning post-training framework that fine-tunes pretrained Action Chunking Transformer (ACT) policies for precision industrial contact manipulation tasks. The method reformulates policy optimization at the chunk level, introduces an actor-critic architecture adapted for ACT, and uses a hybrid behavior-prior constraint to prevent distribution shift during online RL fine-tuning. Experiments on industrial contact benchmarks show significant improvements in task success, contact stability, and force safety — including a 46x reduction in force readings above 60 N on a contour-following task — while preserving low latency and GPU memory usage.
RRFP: A Readiness-Driven Runtime for Pipeline-Parallel Training Under Runtime Variability
The paper introduces Runtime-Readiness-First Pipeline (RRFP), a new runtime for pipeline-parallel large-model training that treats schedules as non-binding hint orders rather than strict execution sequences. By combining message-driven asynchronous communication, lightweight tensor-parallel coordination, and ready-set arbitration, RRFP dynamically dispatches work based on actual task readiness, reducing idle bubbles and stage misalignment. Implemented on a Megatron-based framework and evaluated at up to 128 GPUs, RRFP achieves up to 1.77× speedup on language-only workloads and 2.77× on multimodal workloads versus fixed-order baselines, and outperforms the fastest comparable external system by up to 1.84×.
RELEX: Extrapolating LLM RLVR Training via Rank-1 Parameter Trajectories
This paper demonstrates that RLVR weight update trajectories are extremely low-rank and near-linearly predictable, with a rank-1 approximation capturing most downstream performance gains. The authors propose RELEX, a compute-efficient method that observes a short training window, estimates the rank-1 subspace, and extrapolates future checkpoints via linear regression—requiring no additional training. Evaluated on Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base, RELEX matches or exceeds full RLVR performance using as few as 15% of training steps, and can extrapolate up to 10–20× beyond the observed prefix. The authors attribute the method's effectiveness to a denoising effect from rank-1 projection that discards stochastic optimization noise.