Preliminary: pilots, not a leaderboard.

Unitree G1 table tennis pilot. Robotics Engineer mode: the agent gets the robot model, an IK library, a joint-target interface with real limits and Unitree's walking policy, and must implement hit(t, pos, normal, vel) itself. After the match the final implementation is scored in a separate container on 29 hidden, feasible hit requests (success = position ≤ 8 cm, face ≤ 10 deg, speed error ≤ 0.5 m/s + 30%, time ±20 ms, no rejected command, no fall). Observation: true state (no cameras) in this pilot. Slowed ×10, L1 seed 101.

• Tolerances are calibrated for G1 and are much wider than for Franka.
• 29 hidden requests only; Astra's first run (28/29) is near the ceiling.
• Astra's third run scored 12 of 29 hidden requests.

Epee fencing, Franka Panda + 1-axis rail, against scripted L1. Skill API strike(t, point, dir, vel). Robotics Engineer mode: the agent implements strike from joint interface + IK library; post-match API test on 40 hidden requests and a competition test (our fixed reference strategy drives the agent's strike, 24 games).

• L1 is too weak to separate models in the Athlete mode (all won).

Table tennis, single player against the scripted opponent (L1 seeds 101 and 111, L2 seed 202), Athlete mode, cameras only, slowed ×10. In the logging condition the player's program writes its own estimate (ball position and velocity, predicted bounce and contact point, opponent paddle position) before every command; after the match the recording is replayed bit-exactly and each line is compared with the simulator state.

• This measures perception and physical prediction, not reading the opponent.
• Asking for the log changes the program (SOL and Astra started aiming away from the opponent once asked to report its position); run both conditions.
• Only 3 matches per model per condition.

Table tennis duels GPT-6.1 SOL vs GPT-6 Astra; the between-point reviews are scored by an LLM judge for opponent-directed understanding or prediction; conditions: no prompt / prompt asking to read the game / scouting report, 5- and 11-point games.

• LLM judge (GPT-6 Astra after the fact-label fix); the judge is lenient.
• Prompting is itself an intervention; this is a diagnostic, not a measure of spontaneous game reading.

Table tennis single player against the scripted opponent (L1 seed 101, L2 seed 202), slowed ×20, cameras only, both modes; each body evaluated as a prototype. One match per body × mode × level × model.

• Research variant: separate from the standard results table (one fixed body per sport in the Athlete mode).
• One match per cell; differences between models are smaller than differences between bodies.