# Embodied Agents

Embodied agents are asked to work like students in a teaching laboratory: set up the apparatus, choose the conditions, read the instruments and report a physical quantity. Prior manipulation benchmarks only scored reaching a given target state, while scientific-discovery benchmarks replaced experiments with text commands or numerical queries. EmbodiedLab sets 30 quantitative experiments across five disciplines on 14 apparatus platforms, resamples each instance's hidden parameters, and scores the result together with the readings recorded in that episode and the restored bench; run in their developers' coding agents, Claude Opus 5.5 reaches 43.3% success under native control and GPT-6 Astra 40.0%.

Embodied agents are asked to work like students in a teaching laboratory: set up the apparatus, choose the conditions, read the instruments and report a physical quantity. Prior manipulation benchmarks only scored reaching a given target state, while scientific-discovery benchmarks replaced experiments with text commands or numerical queries. EmbodiedLab sets 30 quantitative experiments across five disciplines on 14 apparatus platforms, resamples each instance's hidden parameters, and scores the result together with the readings recorded in that episode and the restored bench; run in their developers' coding agents, Claude Opus 5.5 reaches 43.3% success under native control and GPT-6 Astra 40.0%.

RoboFoundry proposes a Self-Evolving System-as-Policy framework that treats the entire supporting system around a foundation model—its context system and hierarchical skill system—rather than model parameters as the policy to evolve, diagnosing capability gaps from execution traces, converting them into validated task-specific system updates, and promoting recurring improvements into general system capabilities; it achieves state-of-the-art results on EmbodiedBench (improving GPT-5.5 by 27.8% and bringing Qwen3.7-Plus to 70.3% versus GPT-5.5's 72.7%), outperforms all baselines on RoboMemArena long-horizon memory by at least 39.0%, beats Cap-Agent0 on LIBERO-PRO by 243.8%–679.7%, and demonstrates zero-shot transfer and online evolution on real robots.

RoboFoundry proposes a Self-Evolving System-as-Policy framework that treats the entire supporting system around a foundation model—its context system and hierarchical skill system—rather than model parameters as the policy to evolve, diagnosing capability gaps from execution traces, converting them into validated task-specific system updates, and promoting recurring improvements into general system capabilities; it achieves state-of-the-art results on EmbodiedBench (improving GPT-5.5 by 27.8% and bringing Qwen3.7-Plus to 70.3% versus GPT-5.5's 72.7%), outperforms all baselines on RoboMemArena long-horizon memory by at least 39.0%, beats Cap-Agent0 on LIBERO-PRO by 243.8%–679.7%, and demonstrates zero-shot transfer and online evolution on real robots.

The work introduces HarnessPAI, a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable and evolvable interface organizing the underlying action primitive: within a rollout it executes open-loop with a fixed program, and across rollouts it evolves closed-loop by using execution feedback to revise the program and distill failures into reusable skills; across desktop robot arms, household robots, a robot vacuum, and a legged walking agent it improves on both pure action models and code-as-policy baselines without retraining the underlying model, with a 61.6-point gain over π0.5 on LIBERO-PRO and a 27.2-point gain over WorldDreamer on RoboCasa atomic tasks, and the converged program also serves as an expert-data collector whose data lifts π0.

The work introduces HarnessPAI, a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable and evolvable interface organizing the underlying action primitive: within a rollout it executes open-loop with a fixed program, and across rollouts it evolves closed-loop by using execution feedback to revise the program and distill failures into reusable skills; across desktop robot arms, household robots, a robot vacuum, and a legged walking agent it improves on both pure action models and code-as-policy baselines without retraining the underlying model, with a 61.6-point gain over π0.5 on LIBERO-PRO and a 27.2-point gain over WorldDreamer on RoboCasa atomic tasks, and the converged program also serves as an expert-data collector whose data lifts π0.