RLE-Bench/kinex-benchmark-native-control-budget

Kinex Benchmark (Native Control Budget)

Current campaign: three selected Codex evaluations with GPT-6 Astra / high, seed 0 and the existing native control budgets. Completed results replace their selected task records. Formal statistics and usage are derived from the current selected records; unrelated results retain their existing evidence.

Archived campaign descriptions

Current campaign: eight selected Kinex evaluations with GPT-6 Astra / high, seed 0 and the existing native control budgets. Completed results replace their selected task records. Formal statistics and usage are derived from the current selected records; unrelated results retain their existing evidence.

Archived campaign descriptions

Current campaign: GPT-6 Astra / high across 40 LIBERO, 50 RoboTwin, and 42 standard RoboDojo tasks for both Codex and Kinex. This continuation adds 132 Codex and 117 Kinex episodes, retaining the existing first five Kinex results per simulator from the same frozen sources. Current catalog instructions, authored per-task seeds, and native control budgets apply. Random RoboDojo is excluded. No automatic episode retries. Verified results are published periodically; unrelated results and their artifacts are retained.

Archived campaign descriptions

Current campaign: fifteen fresh Kinex episodes with GPT-6 Astra / high: task02/01-05, task03/01-05 and task04/01-05. Seed 0 and current catalog instructions. LIBERO uses 605 raw steps (600 agent action steps after five initialization steps); RoboTwin uses the official task integers 800, 600, 1700, 500 and 1200 under unchanged RoboEnv execution; RoboDojo uses 1875 steps. Full-suite, Codex and random RoboDojo evaluations are excluded. No automatic episode replacements. Completed results are published periodically. Previous Astra results are archived with their original measured budgets; all other models and historical artifacts are retained.

Archived campaign descriptions

Current defaults for new evaluations

Environment	Public action-step budget	Control time
LIBERO (all four suites)	600	30 s
RoboTwin	Official per-task 400–1,700	16–68 s at 25 Hz
RoboDojo	1,875	75 s at 25 Hz
SIMPLE	3,750	75 s at 50 Hz
RoboCasa	6,000	300 s at 20 Hz

LIBERO action-step usage and limits on the site and in CSV exports exclude five environment initialization steps. Its runtime limit is 605; raw verifier counts and measured protocols retain initialization. CSV exports also retain the raw counters. Historical episodes are displayed against their own recorded limits, and frozen campaign plans below are preserved. RoboTwin uses the official integer limits with the existing RoboEnv execution and counting unchanged.

Current definitions: benchmark source. These defaults do not start or resume evaluations or change existing packages.

Retained campaign records

Current campaign: nine fresh Kinex episodes with GPT-6 Astra / high: task02/01,02,04,05,07,10; task03/04; task04/01,03. Seed 0 and current catalog instructions. Budgets: LIBERO 525, RoboTwin 2500, RoboDojo 1875 native control steps. Full-suite and Codex evaluations require a later session. No automatic episode replacements. Completed results are published periodically. Previous Astra results are archived; all other models and historical artifacts are retained.

Archived campaign descriptions

The previous GPT-6 Astra / high campaign reruns both Codex and Kinex on LIBERO (40), RoboTwin (50), and the 42 standard RoboDojo tasks: 264 fresh task-agent episodes. It uses updated catalog instructions and their public scene bindings. RoboDojo receives 1875 native control steps per episode; LIBERO and RoboTwin use the recorded task-specific budgets. Variant-specific planned budgets prevent changes to retained historical configurations. The 12 random RoboDojo variants are excluded. A valid native failure is retained; after individual review and repair, a provider or infrastructure interruption may receive at most one fresh replacement. All attempts and their usage remain available.

Seed 0 is used except for four RoboTwin initialization exceptions: task03/16 uses seed 5, task03/25 uses seed 1, and task03/45 and task03/46 use seed 2. Their seed 0 preflights failed with native unstable-object errors before any model call. These previously disclosed seeds were explicitly approved for this rerun, use the same seed for both agents, and must pass fresh preflight checks. Runtime seed substitution is disabled.

The Opus variants now hold two runs at this revision (see the Opus section); their earlier results (robo-rsi nv1, revision 0163563) remain in attempt-history.json. The Astra boards now contain only this rerun, with verified completed episodes published periodically. All previous Astra episodes remain in attempt-history.json; scheduled tasks also expose them in Previous attempts. Historical records retain their original instructions, budgets, videos and immutable evidence links. New source identities appear in campaign-status.json and each episode. New task budget documentation is at campaigns/astra-native-rerun-20261012T224235Z/task-budgets.md in the Dataset.

Historical Astra campaign

GPT-6 Astra with high reasoning effort and per-task native evaluation control budgets. This campaign evaluates Codex on 164 tasks across RoboCasa (10), LIBERO (40), RoboTwin (50), RoboDojo (42 standard tasks), BEHAVIOR (10), and SIMPLE (12). Kinex is evaluated on the same 42 RoboDojo tasks, for 206 selected episodes total. Other Kinex environments are not scheduled in this campaign. RoboBallet and the 12 random RoboDojo tasks are excluded.

Each pair uses a fresh workspace and one episode, seed 0 by default. Any native initialization seed exception is disclosed and shared across agents on that task. Infrastructure interruptions may receive a fresh replacement attempt; all attempts are retained. Valid native task failures are not retried. Pending tasks are excluded from success denominators. Missing usage remains unavailable, not zero.

Budgets follow the reviewed per-task native evaluation plan. Some horizons are explicit estimates; RoboTwin uses a permissive conversion compatible with the current RoboEnv frame counter. See task-budgets.md in the paired Dataset and each task's planned and measured protocol for exact budgets.

The interface preserves the reference Kinex Benchmark layout: overview, agent reports, task browser, native video, full session, artifacts, evidence and attempt history. Each episode retains the source identity of its own campaign. Sources are pinned in sources.json; campaign-status.json records current progress.

Space: https://huggingface.co/spaces/RLE-Bench/kinex-benchmark-native-control-budget Dataset: https://huggingface.co/datasets/RLE-Bench/kinex-benchmark-native-control-budget

A provider error that occurs only after the agent receives a valid sealed native terminal verdict does not replace that native result. Such episodes retain the original verdict, raw harness interruption, complete reported usage, and an explicit completion review. This rule applies to either native success or native failure; the episode is not rerun.

Token totals retain all identified provider-reported responses, including context compaction. When CLI completed-turn summaries omit explicitly identified compaction responses, exact reconciliation against both those summaries and native cumulative counters is retained in a per-episode usage review. Original token totals and raw evidence remain unchanged; unexplained differences block publication.

Three BEHAVIOR executions continued inside their original containers after their host collectors disappeared during a chat interruption. Replacement collectors observe those same Docker executions and invoke the unchanged native verifier when they end. Per-episode collector reviews verify session continuity, original evidence prefixes, retained session inode bytes and checksums. Model processes and simulators are not restarted. Collector-generated result records and UUIDs are explicitly identified; agent wall time is derived from the original session timestamps, and unavailable host setup timings remain null.

For the three still-running BEHAVIOR episodes (task05/01, /04, /07), the assistant prospectively extended host waiting protection from 8 to 16 hours at 2026-10-12 13:58:54 UTC. Observed frame throughput suggested that two episodes could otherwise hit the host timer before using their assigned native control budgets. This operational adjustment is not an official simulator horizon: native frame budgets, models, seeds, original task configurations and original executions are unchanged. The original requested timeout remains recorded; per-episode CSV and protocol metadata disclose the effective waiting limit. See operational-wall-time-policy.json for the decision and contemporaneous evidence.

Opus 5.5 on Vertex AI (robo-rsi, two runs)

Eight Opus 5.5 variants (1M-token context, Vertex AI) are two runs of the same protocol, Claude Code and Kinex at high and xhigh effort. Run 1 keeps the earlier variant ids (cc-opus-vertex-high, cc-opus-vertex-xhigh, kinex-opus-vertex-high, kinex-opus-vertex-xhigh; model label "Opus 5.5 · Vertex AI · run 1", campaign opus-vertex-run1-20261013T043000Z); run 2 uses the same ids with a -run2 suffix (model label "· run 2", campaign opus-vertex-run2-20261013T043000Z). Both use this benchmark revision (c006b8d) with its native control budgets and reviewed goals on RoboCasa, LIBERO, RoboTwin, RoboDojo (42 standard tasks) and SIMPLE: 154 tasks, one fresh episode each, the manifest's per-task seeds (RoboTwin 16:5, 25:1, 45:2, 46:2, as codex-astra), an eight-hour agent timeout and no reruns of scored episodes. BEHAVIOR and RoboBallet are not included. Sources: RoboEnv a187ff9 (dev/v1.0.0), Kinex e6cc9de (PR #32), Claude Code 2.1.296, Kinex without strict tool schemas (Vertex AI policy). Simulator images carry RTX PRO 6000 Blackwell environment patches (OIDN 2.3.3 for RoboTwin, TAA instead of DLAA for SIMPLE, a RoboDojo camera-readiness hook). Infrastructure failures before scoring ran again.

Vertex AI 30 MB request limit (Kinex, RoboDojo). Kinex keeps every camera image in context, and on long RoboDojo episodes a request can exceed the Vertex AI 30 MB limit (HTTP 400), which ends the agent mid-episode. Such episodes ran again, at most five attempts. For 14 tasks every attempt in one run was cut off this way; their formal result is the same variant's valid episode from the other run (marked substituted_from; listed in the table), so the two runs are not independent on those tasks. All 14 filled results are native failures. The cut-off attempts are in attempt-history.json. Two run-1 kinex-opus-vertex-xhigh episodes (task04/14, task04/53) had already succeeded natively when the next request hit the limit and were rerun in error; their original results are restored with a post_terminal_review (as codex-astra task03/35), and the reruns are in the attempt history.

Usage and cost. Claude Code usage is its native session usage, reconciled with its modelUsage. Kinex usage is every response's usage from the Pi session (Anthropic reports cache reads, cache writes and reasoning on every response); RLE's request-level provider audit supports OpenAI Responses only, so Kinex request attempts are not observed (audit_complete: false), and requests Vertex AI rejected count as rejected_attempts. Costs are API-equivalent at Anthropic list prices for claude-opus-5-5: $4 input, $5 cache write (5-minute; no 1-hour writes occurred), $0.20 cache read, $20 output per MTok; global endpoint (no regional premium) and the 1M context at standard pricing. Earlier results of these variants (robo-rsi nv1, revision 0163563) are in attempt-history.json.

Variant	Run	RoboCasa	LIBERO	RoboTwin	RoboDojo	SIMPLE	Total	Filled from other run
cc-opus-vertex-high	1	1/10	36/40	45/50	17/42	7/12	106/154	—
cc-opus-vertex-high-run2	2	1/10	39/40	45/50	23/42	6/12	114/154	—
cc-opus-vertex-xhigh	1	1/10	40/40	48/50	26/42	9/12	124/154	—
cc-opus-vertex-xhigh-run2	2	1/10	38/40	47/50	25/42	7/12	118/154	—
kinex-opus-vertex-high	1	2/10	37/40	45/50	21/42	7/12	112/154	task04/03, task04/16, task04/17, task04/21
kinex-opus-vertex-high-run2	2	2/10	37/40	47/50	19/42	9/12	114/154	task04/01, task04/06, task04/24
kinex-opus-vertex-xhigh	1	1/10	40/40	46/50	25/42	7/12	119/154	task04/03, task04/24, task04/29
kinex-opus-vertex-xhigh-run2	2	1/10	40/40	47/50	26/42	10/12	124/154	task04/06, task04/16, task04/17, task04/21

The Astra native-budget pilot, main launch, and authorized replacement assets share campaigns/astra-three-family-full-20261013T044142Z. Original launch identifiers, episode IDs, native results, usage, and source provenance are retained. The older protocol experiments and independent Opus runs keep their own campaign directories.