maniskill

🔥 [ICLR 2026] Benchmark for robotic tabletop manipulation memory-intensive tasks

Squint: Fast Visual Reinforcement Learning for Sim-to-Real Robotics [PyTorch, SO-101 Robot Arm, ManiSkill3, Sim-to-Real]

🔥[AAMAS 2026 Oral] Don’t Blind Your VLA: Aligning Visual Representations for OOD Generalization. https://blind-vla-paper.github.io

A minimal Vision-Language-Action model you can read: frozen CLIP + a tiny head on ManiSkill PickCube. LeRobot integration. Runs on a Mac, no GPU.

Minimum example showing how to setup Singularity with Maniskill3 to run training headless on a cluster.

APPL is an agentic robot learning framework connecting policy learning with agent-based skill selection and composition through structural priors for generalization from a few demonstrations to new scenes and task compositions.

Asset files for the Trossen Robotics WidowX AI robot for the ManiSkill framework

Benchmark for adaptation to hidden robot action-interface contracts: 4 ManiSkill tasks, preregistered gates, belief/probe/learned methods, delay-aware control. Same policy, different wiring — can it adapt?

PPO reinforcement learning on the Unitree G1 humanoid — learning bipedal balance in MuJoCo and dexterous apple-in-bowl manipulation in ManiSkill 3, with side-by-side training-progression demos.

Training and deploying ACT policies for SO-101 cube grasping: motion-planning data collection, ManiSkill-to-LeRobot conversion, then simulation and real-robot evaluation.

Reproducible ManiSkill PickCube visual imitation-learning workflow for MLP BC and ACT.

GPU-parallel visual PPO for ManiSkill StackCube: asymmetric Actor-Critic, four-frame dual-camera input, normalization, TCAPS and NoEarlyReset.

Predicted Futures Are Not Enough: Learning Executable Goals for Robot Manipulation — code, frozen goal banks and evaluation records

双相机 Qwen3.5 VLA：可信数据、Rectified Flow Action Chunk、ManiSkill 闭环评估与 temporal ensemble

Tight-clearance insertion demonstrations in ManiSkill: 4cm cube into a 4.2cm tray cell, and a phone into a slot via two-Panda handoff. Collected for PI0.5/VLA training.

ManiSkill PickCube PPO study comparing privileged state, RGB, and RGB-D observations across three training seeds.

Reproducible comparisons of open robotics AI models, data pipelines, simulators, and contribution workflows

Reproducing PPO on ManiSkill's PushCube task, with reward shaping and curriculum learning experiments.

OOD robustness of visual robot manipulation policies in ManiSkill with PPO and domain randomization.

Behavior Cloning lab for ManiSkill PushCube-v1 with PyTorch.