StarVLA-RSI agent instructions

This repository evaluates complete agent methods, not only bare model names, and is not a model-training repository. Participants may compose multiple general-purpose models and own their planning, memory, tools and decision loop. Direct control limits environment permissions and robot execution interfaces, not internal reasoning. Protocol-first integration is primary; Python Policy and the reference loop are optional. Keep method/model roles and configuration explicit; report measured usage without inventing missing attribution.

Preserve native benchmark execution budgets. Existing method/track safety caps remain declared, not universal limits on inference. Interaction/media recording is optional, default enabled, and never a publication prerequisite. Essential results, denominators, counters, configuration and errors survive minimal mode. Retention options must not change actor inputs, controls, lifecycle or scoring; runtime scratch inputs live until episode closure. Internal hashes/source audits remain support tools, separate from public result eligibility.

Action chunk length, repeated controls per decision and replan timing belong to the policy method and its trusted driver, not a universal server-side prefix. The current service accepts one action per sequenced step; the driver may send successive controls from one decision, subject to action validity, native control, termination and the declared total execution budget. Native physics substeps and protocol-fixed interpolation are not policy-adjustable repeats. Legacy decision-budget tracks still couple call/prefix settings to execution caps, and some interfaces declare required_replan_interval; preserve those contracts until an explicit migration. Do not claim all existing tracks are decoupled.

The product scope is direct control by general-purpose models across many benchmarks. Do not add VLA-assisted/hybrid control, learned action proposals, fallback task policies, or automatic task-solving skill services. The model must select the actions through the declared observation-to-decision interface. Only pinned native low-level control and protocol-fixed action conversion may execute them; they must not plan the task for the model. External reports are design references, not permission to expand observations or control capabilities.

Use a multi-agent implementation -> independent review -> regression validation -> correction loop for framework changes. Assign non-overlapping implementation files, review cross-layer effects independently, and test a stable source snapshot before applying changes to the evaluation checkout. A passing unit test is not evidence of native evaluation, provider attestation or complete benchmark parity.

Before changing evaluation code or protocols:

Read docs/architecture.md and docs/evaluation.md. Internal project_readmes/ notes, if present, are local context, not required public docs.
Read the target benchmark's EVALUATION_PROTOCOL.md and its native evaluator.
Treat the pinned upstream evaluator as the source of truth. Do not guess action semantics, success criteria, episode counts, seeds, or observation layouts.
Record verified evidence separately from migration TODOs.

New benchmark integrations must use the contracts under benchkit/, delegate task success to the native benchmark, and add a dependency-light test. Training recipes and dataset registries do not belong under benchmarks/simBenchmarks.

Evaluation artifacts must identify the benchmark and protocol versions, policy and model, seed and episode, action contract, source revision, token/cost/latency usage when available, and all errors. Infrastructure errors must not be silently counted as policy failures.

The primary control architecture has two ends, not a mandatory harness middle layer. Model-specific trusted evaluation entrypoints belong under AgentPolicies/<profile>/; reusable agent-side decision loops belong under AgentPolicies/policy_support/. Benchmark adapters, native runtime/asset helpers, task inventories and evaluation plans belong beside their protocols under benchmarks/simBenchmarks/<benchmark>/ (future hardware adapters under benchmarks/realRobots/); shared SIM transport lives in benchkit/sim_server/, which owns reusable communication and lifecycle mechanics, not model decisions or native task semantics. benchkit/ also supplies result/data formats, recording, source capture, audits and reporting tools. Fixed-plan assembly belongs in benchkit/campaigns, report import/coverage tools in benchkit/reports, and benchmark metadata in benchkit/catalog. Relocated modules do not keep old-entrypoint forwarding shells. Import the optional runner directly from AgentPolicies.policy_support.runner. Do not rewrite frozen old packages. Benchmark code owns task selection and native success; the model owns decisions. Central task scheduling is retained in this version: coordinating fixed plans, case claims, launches and progress is distinct from making robot decisions. Current campaign commands use their frozen local execution path; do not describe them as remote-service orchestration or distributed recovery. Preserve the single-attempt and uncertain-ACK/no-replay rules when extending scheduling. Plugin discovery, a comprehensive unified CLI and cross-benchmark leaderboards are long-term goals, not requirements to build in every cleanup iteration. An explicitly reviewed trusted agent driver may hold the administrative SIM client for fixed claims and result collection. The model actor must receive only declared observations/counters, never that client, its credentials or management/scoring methods. An administrative client is not actor-safe merely because its file moved to another directory. No native simulator imports in model drivers or policies. Shared fixed-action infrastructure fixtures belong under benchkit/diagnostics/, not AgentPolicies/. They may implement Policy only for explicit non-capability diagnostics; no model inference or task-solving strategy belongs there. Benchmark-specific native readiness tools remain under its environment/.

Model-specific policy code/configuration and evaluation results belong under AgentPolicies/<model-profile>/; shared policy behavior belongs under AgentPolicies/policy_support/. New reports, coverage lists and videos use the model's results/<benchmark>/ namespace. Historical anonymous trials must not be renamed as verified named-model scores. Preserve frozen original campaigns and result bytes; relocation indexes and byte-checked presentation copies are allowed.

User-facing rollout export directory names must include the canonical task ID and a short English task description, in addition to date/round/case identifiers. Do not use only names such as 2026-09-07-round2-case000. For example: 2026-09-07-stack_bowls-bowl-stacking-round2-case000/video_cam_head.mp4. Keep camera identity in video filenames. New video exports go under the model's result namespace, not examples; old examples paths may retain compatibility links. Separate reset-only and no-policy diagnostic exports under rollouts/archive/ and label their purpose. Never rename frozen campaign identities or overwrite original results to tidy exports; record relocations in the rollout README and preserve original artifact bytes.

Local commits are allowed. Never run git push or otherwise change remote Git state.

Persistence and voluntary give-up rules are policy behavior, not benchmark rules. A policy may explicitly configure a minimum number of executed action decisions in its own versioned configuration/prompt; the shared harness only records execution feedback and enforces declared public hard limits. Never silently turn one model's persistence setting into a requirement on all models. Native success/termination, exhausted limits, errors and explicit safety stops must not be delayed to meet a policy's persistence target. Historical runs retain their original frozen implementation and results; an old policy rule neither changes benchmark definitions nor authorizes another attempt.

fast means a preselected subset with one attempt per selected task. full means the native protocol's declared task/variant/seed/episode schedule, not success-seeking retries. Both retain failures and errors. Full presets are not yet implemented for all benchmarks: never label task-only coverage as native full evaluation or guess its episode count. Refactoring never starts an evaluation.