Jiafei Duan — Astra vs Opus 5.5 in MuJoCo Ring Fence sumo, Astra wins in 31 seconds

TL;DR

Jiafei Duan (NUS, ex-Allen AI robotics lead) rigs a MuJoCo Ring Fence where two Unitree G1 humanoids — one driven by GPT-6 Astra, one by Claude Opus 5.5 — trade whole-body joint-target commands every 2 sim-seconds, with the win condition being to push the opponent out of the ring. Astra wins after 31 seconds. The setup is deliberately minimal: state → LLM → joint targets → act, no VLA, no motion planner, no learned policy anywhere in the loop. It’s a self-contained head-to-head evaluation of two frontier LLMs’ physical-intelligence at the whole-body-humanoid control frequency, one week after Anthropic’s Opus 5.5 launch and directly extending the Astra-as-controller thread already on the wiki.

Key claims

Two Unitree G1 humanoids in a MuJoCo ring, one driven by GPT-6 Astra and one by Claude Opus 5.5, with win condition = push opponent out of the ring [tweet body].
Control loop is 2 sim-second cadence: state → LLM → joint targets → act, no VLA, no motion planner, no trained policy [tweet body].
Astra won after 31 sim-seconds [tweet body].

Method

Two Unitree G1 humanoids are placed in a MuJoCo simulated arena bounded by a “ring fence” borrowed from combat-sports staging. On each 2 sim-second tick, the current world state (joint positions, velocities, contact state, opponent pose implicit in the state feed) is packaged into a prompt, sent to the assigned LLM, and the LLM emits joint target commands that MuJoCo applies over the next tick. No intermediate policy — no VLA, no VLM planner over a VLA, no motion planner, no IK layer beyond the joint-target interface itself. The two agents run concurrently against each other. Failure mode = leaving the ring. Only reported rollout is single-run; no repeated trials or per-model win-rate numbers in the tweet.

Results

Astra ends the fight in 31 sim-seconds by forcing Opus 5.5 out of the ring [tweet body].
No aggregate statistics (win rate over trials, energy expended, ring-exit direction, per-tick joint-command sanity metrics) are reported — one-off qualitative demo [tweet body].

Why it’s interesting

Sits squarely on the VLA Models recipe-lever board’s LLM-as-controller row alongside the growing set of Astra/Fable/Claude-drives-a-robot demos: Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK (Astra vs Fable 5.1 on real YAM arms via IK), Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness (Astra doing physical in-context learning from one human video), thijs — Astra paints the Golden Gate Bridge with a real robot arm (Astra painting the Golden Gate Bridge with a real arm), and the Agent as Policy for Robotic Manipulation paper-scale AGP evaluation. What this one adds is a whole-body humanoid control substrate at 29 DoF (G1) rather than the ≤7-DoF tabletop-arm regime the prior demos live in — the exact regime where How Claude Performs on Robotics Tasks reported “no model stands up the 29-DoF G1 humanoid” via direct joint control, so a 31-second successful sumo-push is a fresh data point on how fast that ceiling is moving. The adversarial framing is also novel for the concept: prior LLM-as-controller filings measure success against a fixed task; this one is a symmetric two-agent competition, which turns the eval into a relative ranking rather than an absolute benchmark and sidesteps the usual “did the environment give partial credit” scoring ambiguity — the ring boundary is a hard binary. Sits on RL Environment Platforms as a new eval-substrate cell: MuJoCo + LLM-controlled humanoid competition, distinct from the packaged-task-suite pattern (Isaac Lab RoboLab, RoboCasa365, Genesis) and closer in framing to the Automataleague RL sumo-policy demo the reply thread points at — same arena, learned RL policies vs frontier LLMs as the natural next comparison.