grounded-intelligence-initiative/ale-robotics-task-logs

ALE robotics task logs

Task logs of AI agents on the tasks of the Agentic Robotics Benchmark, run with the benchmark's ALE engine (each episode in its own sandbox, graded by the task's verifier). A task log is one episode of one agent on one task: the agent's full trajectory, the results and the verifier's log, with the agent's submission. The dataset keeps one task log per task and agent: the latest scored episode. When a new episode of the same agent on the same task is scored, it replaces the folder.

42 task logs. Platforms: arm 23, driving 1, hand 3, humanoid 3, quadruped 2, space 2, swarm 1, uav 7. Agents: claude_opus55-claude_code-high 28, kimi31_pro_replica-kimi_code-max 14. Last updated 2026-11-09.

Layout

index.jsonl                      one row per task x agent (fields below)
<platform>/<task>/<agent>/       one task log
  run.json                       run summary + per-step digest (format 3, below)
  README.md                      this task log in brief
  trajectory.json                the ATIF trajectory as the engine wrote it
  trajectory.jsonl.gz            the trajectory, one JSON object per line
  steps.bin                      every distinct picture the agent saw, once (byte ranges in run.json)
  pics/                          pictures that had no file of their own in the records, decoded
  poster.jpg                     a still for the run
  results/                       result.json, verification.json, lock.json (resources.jsonl when sampled)
  verifier/                      stdout.txt, stderr.txt (the grader's log), diagnostic.txt
  stages/                        stdout / stderr of every other stage command, commands.json
  traces/                        trace.execution.jsonl, trace.transport.jsonl
  blobs/                         pictures and large texts the trajectory references
  artifacts/                     the submission the grader scored
  replay.mp4                     a replay of the submission on a development episode, where one exists

• platform: the task's [metadata].platform (arm, hand, humanoid, uav, driving, space, swarm, quadruped, ...).
• task: the task's name in the benchmark (tasks/<platform>/<task>/). A task measured before its publication under an earlier name is filed under its current name (measured_as keeps the old one).
• agent: <model>-<harness>-<effort>, e.g. claude_opus55-claude_code-high = model claude-opus-5-5, harness Claude Code, reasoning effort high (default when the harness's default was used).

Which episode: per task, the version (content digest) of its newest scored episode; per agent, the newest scored episode of that version. An episode lost to an infrastructure failure without a score never replaces a scored one. Only published tasks are included (the benchmark's main branch or an open task pull request).

index.jsonl

field	meaning
platform, task, agent	as above; path is the task log's folder
model, harness, effort	the agent as the engine recorded it
run_id, episode_id	the ALE run and episode
date, started_at	when the episode started (UTC)
task_commit, task_content_digest	the task version that ran: its commit (null when the engine ran a local copy) and the engine's content digest
measured_as	the task's name when it was measured
task_location	where the task lives now: main or the pull request that carries it
status	the engine's status: completed, timeout (the agent's time ran out; its submission was still graded), agent_error, env_error
reward, reward_exact	the score in [0, 1]: rounded half-up to two decimals, and exact
success	1 when the reward reached 1.0
metric, metric_value, metric_direction, anchor, full_at	the task's metric, its measured value, whether higher is better, the reference value and the ratio at which the reward saturates (reward = clamp(clamp(value / anchor, 0, 1.5) / full_at, 0, 1); inverted when lower is better)
cost_usd	model cost at the list price per token type (input, output, cache read, cache write) times the gateway's token counts
tokens_in, tokens_out, model_calls	gateway totals (tokens_in includes cache reads and writes)
agent_time_s	the agent phase, seconds
steps, pictures	model calls in the digest, distinct pictures the agent saw
replay, notes	the replay file if any; notes

run.json

Header: format (3), slug (<task>--<episode hex>), task, bench (the platform), mode (autonomous), run (harness, model, effort), agent (harness), model, effort, state / success, ended_by (infra for an infrastructure failure), reward (every grader metric plus reward, reward_exact, success, infra_error and resource fields, null where not recorded), wall_s, tokens (in, cached, out, usd and the details), task_text (the instruction the agent got), replay, poster, steps_bin, trajectory.

steps[]: one entry per model call: id (its first ATIF step), t (seconds since the agent phase began), say (the agent's text, in full), calls[] (kind: shell, edit, read, look, robot = the task's own simulator, robot, CAD or practice tools, web, other; label; code, long code cut in the middle), out (text, cut in the middle when long; the full text is in the trajectory; exit; pics = [offset, length, mime] ranges of steps.bin), tok (that call's tokens and USD).

Also: ale (run and episode, status, failure, engine commit and version, harness version, phases with timings), score (metric, direction, measured value, anchor, full_at, the rule, the grader's diagnostic and criteria, the grader log's tail), task_info (the measured version: commit, content digest, name; location: where the task lives now and its latest commit), tasklog (layout version, path, the file map, the verifier files, files left out), files, replays, poster_info, resources (limits; peaks when sampled), digest (how steps were built), pictures (every picture: sha256, size, mime, file, byte range, the steps that show it), provenance (lock.json).

Formats

• ATIF (trajectory.json): the Agent Trajectory Interchange Format (Harbor ATIF v1.x): steps with the agent's messages and exposed reasoning, tool calls and their results; pictures and large outputs are files in blobs/ referenced by relative path. When the engine could not convert the harness's native log, trajectory.json keeps only the instruction and the final message and trajectory.jsonl.gz is the native session log verbatim (run.json digest.source says which).
• Traces: trace.execution.jsonl (engine events: phases, commands, their stdout / stderr inline or as files) and trace.transport.jsonl (one event per model call: model, token counts, cache reads and writes, latency, retries).
• Results: result.json (status, rewards, phase timings, the native session id), verification.json (the grader's output), lock.json (task, image, engine, agent and limits).

Host-specific details (local paths, gateway addresses) in the engine's records are replaced by placeholders such as <local> and <gateway>; the agents' own logs are as recorded. Model credentials were never exposed to the agents (lock.json credential_exposed_to_agent: false).

Task logs

platform	task	agent	reward	status	date
arm	arm_hidden_mechanism_box	claude_opus55-claude_code-high	0.09	completed	2026-10-29
arm	arm_hidden_mechanism_box	kimi31_pro_replica-kimi_code-max	0.03	completed	2026-11-08
arm	bimanual_cad_design_from_photos	claude_opus55-claude_code-high	0.0	completed	2026-10-29
arm	bimanual_cad_design_from_photos	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
arm	bimanual_requirements_replica_from_photos	claude_opus55-claude_code-high	0.47	completed	2026-10-29
arm	bimanual_requirements_replica_from_photos	kimi31_pro_replica-kimi_code-max	0.12	completed	2026-11-08
arm	novel_object_6dof_grasp_detection	claude_opus55-claude_code-high	0.17	completed	2026-11-07
arm	object_finding_and_organizing_on_messy_desk	claude_opus55-claude_code-high	0.02	completed	2026-11-08
arm	pump_skid_startup_and_operation	claude_opus55-claude_code-high	0.0	completed	2026-11-09
arm	robocasa_object_dynamics_from_video	claude_opus55-claude_code-high	0.28	completed	2026-10-29
arm	robocasa_object_dynamics_from_video	kimi31_pro_replica-kimi_code-max	0.26	completed	2026-11-08
arm	robotic_arm_cad_design_from_photos	claude_opus55-claude_code-high	0.0	completed	2026-10-20
arm	robotic_arm_cad_design_from_photos	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
arm	robotic_arm_environment_twin_from_real_photo	claude_opus55-claude_code-high	0.17	completed	2026-11-08
arm	robotic_arm_full_assembly_openarm_panda	claude_opus55-claude_code-high	0.02	completed	2026-10-30
arm	robotic_arm_full_assembly_openarm_panda	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
arm	robotic_arm_full_assembly_so101_panda	claude_opus55-claude_code-high	0.0	completed	2026-10-23
arm	robotic_arm_full_assembly_so101_panda	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
arm	robotic_arm_requirements_replica_from_photos	claude_opus55-claude_code-high	0.29	completed	2026-10-21
arm	robotic_arm_requirements_replica_from_photos	kimi31_pro_replica-kimi_code-max	0.16	completed	2026-11-08
arm	robotic_arm_robot_description_from_photos	claude_opus55-claude_code-high	0.1	completed	2026-11-07
arm	shoe_lacing_and_boxing	claude_opus55-claude_code-high	0.11	completed	2026-11-08
arm	shoe_lacing_and_boxing	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
driving	driving_safety_supervisor	claude_opus55-claude_code-high	0.38	completed	2026-11-06
hand	dexhand_shape_sorter_all_pieces	claude_opus55-claude_code-high	0.05	completed	2026-11-01
hand	dexhand_shape_sorter_all_pieces	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
hand	hot3d_sharpa_retarget	claude_opus55-claude_code-high	0.28	completed	2026-11-01
humanoid	furniture_assembly_from_drawings	claude_opus55-claude_code-high	0.11	completed	2026-11-04
humanoid	furniture_assembly_from_drawings	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
humanoid	house_cleaning_with_tools	claude_opus55-claude_code-high	0.0	completed	2026-11-09
quadruped	quadruped_full_assembly_12dof_ur10e	claude_opus55-claude_code-high	0.07	completed	2026-11-09
quadruped	quadruped_requirements_replica_from_photos	claude_opus55-claude_code-high	0.38	completed	2026-11-09
space	fault_adaptive_spacecraft_docking	claude_opus55-claude_code-high	0.06	timeout	2026-11-07
space	fault_adaptive_spacecraft_docking	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08
swarm	hazard_bay_build	claude_opus55-claude_code-high	0.4	completed	2026-11-07
uav	aerial_search_geolocation	claude_opus55-claude_code-high	0.01	completed	2026-11-06
uav	uav_assembly_agipix_xarm7	claude_opus55-claude_code-high	0.0	timeout	2026-11-09
uav	uav_cad_design_from_photos	claude_opus55-claude_code-high	0.11	completed	2026-11-08
uav	uav_requirements_replica_from_photos	claude_opus55-claude_code-high	0.16	completed	2026-11-07
uav	uav_requirements_replica_from_photos	kimi31_pro_replica-kimi_code-max	0.18	completed	2026-11-08
uav	visual_low_level_flight_control	claude_opus55-claude_code-high	0.0	completed	2026-10-31
uav	visual_low_level_flight_control	kimi31_pro_replica-kimi_code-max	0.0	completed	2026-11-08