VSArena V1 walkthrough: platform evaluates embodied/VLA agents in interactive 3D sim without physical robot hardware

A developer released VSArena, an open benchmark evaluating AI agents inside interactive 3D environments where they must perceive, reason, and act rather than only handle text or code. It targets practitioners in VLA, embodied AI, robotics, and agent evaluation, with a public GitHub repo and website soliciting feedback.

Releases·Tool Release

VSArena v0.6.0 launches browser-based Studio for running and inspecting VLA and embodied AI policies

A developer released VSArena v0.6.0, a major update to the browser-based Studio for VSArena, an open evaluation arena for Vision-Language-Action and embodied AI policies built on browser-native 3D physics. It lets users run a policy, watch the environment, and measure results without a local robotics simulator or physical robot.

A developer released VSArena v0.6.0, a major update to the browser-based Studio for VSArena, an open evaluation arena for Vision-Language-Action and embodied AI policies built on browser-native 3D physics. It lets users run a policy, watch the environment, and measure results without a local robotics simulator or physical robot.

Products·launched

VSArena: hosted ELO harness for embodied AI policies

VSArena is an open, browser-based arena for evaluating embodied AI / VLA policies on a single task: a 4-DOF arm stacking three cubes. It provides a hosted harness that runs policies and writes public ELO scores, while physics runs client-side for spectator viewing. It solves the problem of standardized, accessible evaluation for embodied AI.

VSArena is an open, browser-based arena for evaluating embodied AI / VLA policies on a single task: a 4-DOF arm stacking three cubes. It provides a hosted harness that runs policies and writes public ELO scores, while physics runs client-side for spectator viewing. It solves the problem of standardized, accessible evaluation for embodied AI.

Releases·Tool Release

VSArena adds VLA track with camera-only observations, public repo and docs

VSArena, a benchmark for vision-language-action models, now includes a dedicated VLA track where policies receive only a 128x128 RGB camera and a language stacking instruction, with cube poses hidden from the policy. The state-based privileged track is kept separate for debugging and does not affect public ELO. The repository and documentation are now public.

VSArena, a benchmark for vision-language-action models, now includes a dedicated VLA track where policies receive only a 128x128 RGB camera and a language stacking instruction, with cube poses hidden from the policy. The state-based privileged track is kept separate for debugging and does not affect public ELO. The repository and documentation are now public.

Products·building

Open browser-based arena for embodied AI benchmarking

This project is an open, browser-based arena for embodied AI, where Vision-Language-Action models and robotic policies compete on real-time physical reasoning tasks. It aims to provide a transparent, comparable benchmark for embodied AI, similar to LMArena for LLMs, addressing the lack of standardized evaluation in robotics. The demo shows a baseline agent stacking blocks in a client-side physics simulation.

This project is an open, browser-based arena for embodied AI, where Vision-Language-Action models and robotic policies compete on real-time physical reasoning tasks. It aims to provide a transparent, comparable benchmark for embodied AI, similar to LMArena for LLMs, addressing the lack of standardized evaluation in robotics. The demo shows a baseline agent stacking blocks in a client-side physics simulation.