RoboTwin 2.0
PulseAugur coverage of RoboTwin 2.0 — every cluster mentioning RoboTwin 2.0 across labs, papers, and developer communities, ranked by signal.
- developed Fast-WAM 90%
- developed by Fast-WAM 90%
- instance of RoboDojo 90%
- instance of Wamser 90%
- instance of Fast-WAM 90%
- used by V-JEPA 2.1 90%
- competes with GPT -6 Astra 80%
- instance of LIBERO-Plus 70%
- used by World-Action Models 70%
- competes with LIBERO-Plus 70%
- instance of RoboCasa365 70%
- instance of V-JEPA 2.1 70%
16 day(s) with sentiment data
-
Meituan-Robotics-0: New Vision-Language-Action Model for Desktop Manipulation
Researchers have introduced Meituan-Robotics-0 (MR0), a novel vision-language-action foundation model designed for desktop manipulation tasks. MR0 integrates a vast dataset of vision-language samples with extensive robo…
-
La-JEPA: Compact Robot Policies with Latent World Modeling
Researchers have introduced La-JEPA, a novel world-action model designed for robotics that utilizes a compact policy transformer trained from scratch within the feature space of a pretrained V-JEPA 2.1 encoder. This app…
-
New embodied AI models advance robot perception and video generation · 5 sources tracked
Researchers have introduced several new models for embodied AI and video generation, focusing on improving how robots perceive and interact with their environment. AuroraWorld aims to synthesize egocentric videos from e…
-
New REA runtime unifies fast and slow policies for embodied agents
Researchers have developed REA, a novel runtime environment designed to integrate high-level reasoning from foundation models with fast, reactive policies for embodied agents. REA addresses the latency and generalizatio…
-
OmniBrain-0: Unified Embodied Agent for Robots Unveiled
Researchers have introduced OmniBrain-0, a new unified embodied agent designed to coordinate manipulation and navigation tasks for robots. This lightweight model integrates task understanding, planning, progress trackin…
-
New framework EXE extracts transferable manipulation skills from human videos for robots
Researchers have developed a new pretraining framework called EXE, designed to bridge the gap between human manipulation videos and robot learning. This framework extracts shared action representations from egocentric h…
-
POD-RoPE enhances embodied AI with geometry-aware attention
Researchers have introduced POD-RoPE, a novel position encoding method designed for embodied AI systems. This technique represents token observation footprints as distributions, enabling better capture of physical geome…
-
PDRVLA model enhances robot action generation with physical dynamics module
Researchers have developed PDRVLA, a novel vision-language-action model designed to enhance robot control by incorporating a physical dynamics module. This module predicts future visual features, enabling more accurate …
-
Martingale GRPO enhances VLA reinforcement learning by optimizing action penalties
Researchers have developed Martingale GRPO (MART-GRPO), a novel reinforcement learning technique designed to improve the efficiency of training vision-language models (VLMs). Unlike traditional GRPO methods that assign …
-
New World Action Models enhance robot task generalization and real-time action generation
Three new papers introduce advancements in World Action Models (WAMs), a framework for robots to learn latent variables for predicting scene evolution and generating actions. CausalWAM focuses on isolating action-releva…
-
New framework RynnVLA-Latent scales robotic manipulation models using latent actions
Researchers have developed RynnVLA-Latent, a new framework designed to scale the pretraining of vision-language-action (VLA) models for robotic manipulation. This approach utilizes latent actions derived from diverse vi…
-
New research explores context inference in randomized dynamics and dynamic view routing for robots
A new research paper explores how dynamics randomization in simulated environments can affect a model's ability to infer context. The study introduces the concept of "predictive headroom" to quantify the remaining infor…
-
New robot control method DirectWAM slashes inference time by 28x
Researchers have developed DirectWAM, a new method for robot control that significantly reduces inference latency while maintaining high task success rates. This approach converts pretrained video and action experts int…
-
New research advances robot world models for better control and generalization · 8 sources tracked
Researchers are developing advanced world action models (WAMs) for robotics, focusing on improving robot control, generalization, and efficiency. SkeLAM decomposes latent actions into kinematic and residual components f…
-
VLA-Rolling framework reuses discarded actions for faster, more accurate VLA models
Researchers have developed VLA-Rolling, a new framework that improves the efficiency and control of vision-language-action (VLA) models. This method reuses discarded action tails from previous planning cycles, rather th…
-
LoopWAM model improves robot policies using mixed-quality deployment data
Researchers have developed LoopWAM, a closed-loop world-action model designed to improve generalist robot policies by learning from mixed-quality deployment rollouts. Unlike methods that filter out failures, LoopWAM uti…
-
New verification method enhances VLA policy action execution
Researchers have developed a new method called a Policy-Aligned Predictive Verifier (PAV) to improve the execution of action sequences generated by Vision-Language-Action (VLA) policies. This PAV anticipates the outcome…
-
UniW(A)M framework unifies world and action models using diffusion Transformers
Researchers have introduced UniW(A)M, a novel mixture-of-experts framework designed to unify world models and world-action models. This framework utilizes diffusion Transformers to share common physical dynamics while a…
-
RegisterWAM enhances robot action conditioning with novel register approach
Researchers have developed RegisterWAM, a novel approach to World Action Models (WAMs) that enhances robot action conditioning by incorporating learnable non-spatial registers. This method, when combined with observatio…
-
GreenWAM: A Unified Foundation Model for Robotics Control and Prediction
Researchers have introduced GreenWAM, a novel world foundation model designed for robotics. This single model functions as both an action-conditioned world model and a policy, capable of operating across 20 diverse robo…