RoboBridge: A Self-Evolving Embodied Agent Framework for Sim-to-Real Transfer
Preprint. RoboBridge is a framework that represents manipulation tasks as editable procedural skills and wraps a frozen vision-language-action policy as a guided action tool, evolving those skills from execution feedback. On LIBERO-PRO it reaches 74.4% overall success, beating the strongest reported baseline at 72.1%; on three real-robot tasks, 10 episodes of adaptation lift success from 9/15, 7/15, 7/15 to 15/15, 14/15, 15/15. This shifts sim-to-real transfer from retraining to selective skill revision, cutting adaptation iterations by 25% and token use by about 75%.
Whether an agent can preserve reusable procedural task knowledge across simulation and reality and continually revise it from target-environment feedback, instead of relying on policy retraining.
Existing end-to-end VLA policies embed task knowledge in learned parameters, so adapting to real visual and dynamical shifts requires simulator calibration, additional target-domain demonstrations, and further training; tool-using agents make procedures explicit but mostly support execution and reuse within a single environment. Neither provides a clean way to transfer a procedure and selectively revise only the parts that fail in the target environment.
End-to-end models such as RT-1, RT-2, OpenVLA, Octo, and pi0 are adapted through domain randomization, dynamics perturbation, posterior-based simulator tuning, target demonstrations, or fine-tuning; these keep adaptation tied to the policy. Embodied agent systems such as SayCan, Code as Policies, RATs, and ASPIRE add explicit planning, memory, and feedback repair, but they focus on workflow execution or in-domain skill reuse rather than cross-environment procedural migration.
RoboBridge stores each task as an editable skill connecting task intent, observations, tool operations, and outcome verification; a high-level planner selects one operation per turn and receives updated observations and feedback. Candidate skill edits—add, replace, or delete—are evaluated against the current skill and persisted only if validated. A frozen pretrained VLA policy is exposed as an action tool; when needed, the agent builds a subtask reward and applies reward-guided gradients to candidate action sequences during diffusion denoising, without retraining. Simulation and real-robot environments share aligned task semantics and observation/tool interfaces, and skills evolved in simulation initialize real-world adaptation.
On LIBERO-PRO, RoboBridge with Codex/GPT-5.5 achieves 74.4% overall success across SPATIAL, OBJECT, GOAL, and LIBERO-10 under instruction-redirection and position-swap tests, versus 72.1% for Harness VLA and 50.0% for the underlying pi_RLinf policy; it beats Harness VLA by 5 points on GOAL-T. Ablations on SPATIAL-T and OBJECT-T show full framework scores of 83% and 94%; removing skill guidance drops them to 21% and 16%, append-only updates drops to 44% and 57%, removing validation drops to 69% and 63%, and disabling inference-time guidance drops to 75% and 82%. On three real-robot tasks, direct transfer of simulation skills gives 9/15, 7/15, 7/15 versus 4/15, 2/15, 3/15 without skills; after 10 episodes of real-robot evolution, success reaches 15/15, 14/15, 15/15. In an efficiency test on one task, simulation-initialized evolution reaches 15/15 in 12 iterations versus 16 from scratch, and uses 13.08M versus 52.39M tokens, about 75% less.
The authors note evaluation uses only Codex with GPT-5.5 as the coding agent, leaving broader model/harness generalization untested. The real-robot study involves only three tasks and 15 trials per condition, on tasks deliberately aligned with simulation semantics and interfaces; transfer outside that alignment is not shown. The framework relies on proprietary LLMs and multiple pretrained VLA checkpoints, which may introduce cost and reproducibility issues. The paper does not report robot hardware details or code release, and the evaluation does not cover long-horizon safety or ambiguous real-world feedback.
Robotics companies building manipulation systems, warehouse and logistics automation vendors, and general-purpose robot software stacks could use this to shorten sim-to-real deployment and enable post-deployment skill refinement. The approach suits vendors that already maintain a pretrained VLA policy and a structured agent interface. Research labs can apply it now; production pilots are likely 1–3 years away, provided teams can align observation and tool interfaces across sim and real robots, accept LLM inference cost, and add enough validation for safety-critical operations.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.