Tag
Introduces digital teleoperation using action-conditioned world models to generate diverse training data for robotics, decoupling data collection from physical hardware. The system achieves real-time generation and enables zero-shot Sim2Real transfer.
SIEVE is a structure-aware data selection method for vision-language-action imitation learning that identifies reusable visuo-motor primitives and transition interfaces, outperforming full-data training with only 50% of demonstrations and training steps.
Introduces Φ-Nav, a unified on-policy framework that uses hindsight reasoning to synthetically generate path-level instructions from exploratory trajectories, bridging the semantic supervision gap in Vision-Language Navigation and achieving competitive results on R2R-CE and RxR-CE benchmarks with fewer expert demonstrations.
Active-GRPO introduces an adaptive imitation and self-improving reasoning framework that dynamically decides when to imitate references and when to reinforce the model's own discoveries for molecular optimization, achieving statistically significant improvements over previous methods on the TOMG-Bench-MolOpt benchmark.
This paper proposes a noisy expert model to explain the gap between offline and online imitation learning, showing that offline learning from noisy trajectories requires exponential sample complexity while online on-policy distillation achieves polynomial dependence. The analysis leads to an alternative loss function and experiments confirm the theoretical findings.
ReGuide introduces a self-improving framework for diffusion policies that uses test-time guidance to generate corrective rollouts, then fine-tunes the policy on this data, achieving 1.3–7.7× success improvement on Robomimic tasks.
Lerobot SO-101 is a framework that transforms robot learning from messy lab magic into a clean, reproducible software pipeline, enabling structured data collection, policy training, and deployment, similar to what PyTorch did for deep learning.
This paper introduces GRID, a social learning method that extracts universal behaviors from heterogeneous agents by decomposing per-agent reward functions into general and specific rewards, enabling a generalist agent that avoids mode-averaging bias.
This paper investigates whether per-phase metric selection improves demonstration curation for behavior cloning policies. The authors find that phase-gated curation never outperforms global or uniform metric application, and the dilution of defect signals across phases explains the failure.
RL4IL introduces a reinforcement learning-guided retrieval method that uses soft fusion over frozen demonstration libraries to handle missing sensor modalities in robotic imitation learning at inference time, achieving high success rates under complete camera dropout.
ETH Zurich has open-sourced their entire 2026 robot learning course, including slides, lecture recordings, coding assignments, and a GitHub repository, covering topics from imitation learning to foundation models for robotics, with guest lectures from industry leaders.
Researchers propose APT, a two-stage training method that pretrains action experts on vision-action pairs before integrating language conditioning, significantly improving out-of-distribution instruction generalization for Vision-Language-Action policies.
QGF is an RL algorithm that improves policies at test time by using a value gradient to guide a pre-trained flow policy, avoiding training-time instability while maintaining competitive performance.
Introduces StereoPolicy, a framework that leverages synchronized stereo image pairs to improve geometric reasoning for robot manipulation policies, avoiding the fragility of RGB-D and point clouds. It integrates with diffusion-based and vision-language-action policies, showing consistent improvements in simulation and real-world tasks.
Introduces DistIL, a method for reinforcement learning from rich feedback that guarantees monotonic policy improvement, outperforming existing methods on science reasoning, coding, and mathematical reasoning.
Proposes BISON, a system combining learned low-level neural policies with high-level symbolic planning for long-horizon embodied tasks, showing strong generalization and efficiency.
IntentVLA is a history-conditioned visual-language-action framework that improves robot imitation learning stability by encoding short-horizon intents from visual observations, addressing challenges from partial observability and ambiguous observations. It also introduces AliasBench, an ambiguity-aware benchmark for evaluating such methods.
This paper introduces Trust Region Inverse Reinforcement Learning (TRIRL), a method that combines monotonic dual improvement with efficient local policy updates to outperform state-of-the-art imitation learning methods. It addresses the trade-off between stability and computational cost in IRL by using trust-region constraints.
This paper introduces LC-MAPF, a pre-trained model with a learnable communication module for multi-agent pathfinding that improves coordination and outperforms existing learning-based solvers while maintaining scalability.
OpenAI introduced Video PreTraining (VPT), a semi-supervised method that trains neural networks to play Minecraft by learning from 70,000 hours of unlabeled human gameplay video combined with a small labeled dataset. The model learns complex sequential tasks using the native human interface (keyboard and mouse) and demonstrates capabilities like crafting diamond tools and pillar jumping, representing progress toward general computer-using agents.