Dataset Viewer
Auto-converted to Parquet Duplicate
episode_id
string
model
string
task
string
layout_id
int64
frame_first
image
frame_last
image
video_preview
video
model_settings
string
seed
int64
base_task
string
variant
string
category
string
layout_role
string
backend
string
effort
string
outcome
string
score
float64
total_wall_s
float64
official
dict
timing
dict
tokens
dict
tools
dict
agent_messages
list
traces
list
instruction
string
prompt_text
string
videos
list
trace_url
string
s0_l000_align_blocks
gpt-6-astra
align_blocks
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
align_blocks
Open
eval
codex
high
success
1
114.658
{ "take_action_cnt": 109, "action_budget": 200, "ended_by": "official_end" }
{ "start_epoch_s": 1792100429, "end_epoch_s": 1792100544, "total_wall_s": 114.658, "agent_s": 96.154, "sim_s": 18.504, "sim_clock_s": 4.36 }
{ "input": 181066, "cached_input": 154880, "output": 1406, "reasoning": 501 }
{ "calls": 6, "observe": 0, "measure": 3, "run": 3, "step": 0, "assess": 0, "action_status": 0, "errors": 0 }
[ "I’ll push the set square against the three blocks to straighten and square them, keeping the blocks on the table, then return both arms home.\n", "All three blocks now form a straight row, with their edges aligned to the set square. I’ll clear the tool and return both arms home.\n" ]
[ { "step": 0, "item_id": "", "kind": "launch", "tool": "", "arguments_json": "", "text": "\n# This episode\n\nTask: Use the set square to push the three blocks into a straight, aligned row, then reset the robot arm.\n\nThe first observation is already taken, so do not call `goal` or `observe`...
Use the set square to push the three blocks into a straight, aligned row, then reset the robot arm.
You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete the task stated at the end of this message, within the step limit as far as you can; if the steps run short, you must still keep enough of them for the arms' return home. Use only the robot tools. Act autonomously and ne...
[ { "camera": "head", "url": "https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/videos/s0_l000_align_blocks/head.mp4", "bytes": 312693 }, { "camera": "left_wrist", "url": "https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/videos/s0_...
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_align_blocks.jsonl.gz
s0_l000_arrange_largest_number
gpt-6-astra
arrange_largest_number
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
arrange_largest_number
standard
Generalization
eval
codex
high
success
1
355.901
{ "take_action_cnt": 529, "action_budget": 1050, "ended_by": "official_end" }
{ "start_epoch_s": 1792100431, "end_epoch_s": 1792100787, "total_wall_s": 355.901, "agent_s": 237.73, "sim_s": 118.171, "sim_clock_s": 21.16 }
{ "input": 1768687, "cached_input": 1706112, "output": 4111, "reasoning": 1245 }
{ "calls": 36, "observe": 0, "measure": 13, "run": 23, "step": 0, "assess": 0, "action_status": 0, "errors": 1 }
[ "I’ll place the digits as **8720**, upright and centered on the four pads, then return both arms home.\n", "The 8 is centered on the first pad. I recovered the 2 after it slipped and now have a secure grasp.\n", "The first three pads now read 872. I’ll move the 0 across the clear table area so the right arm can...
[ { "step": 0, "item_id": "", "kind": "launch", "tool": "", "arguments_json": "", "text": "\n# This episode\n\nTask: Arrange the numbers from left to right to form the largest possible number, and place them on the pad.\n\nThe first observation is already taken, so do not call `goal` or `obser...
Arrange the numbers from left to right to form the largest possible number, and place them on the pad.
You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete the task stated at the end of this message, within the step limit as far as you can; if the steps run short, you must still keep enough of them for the arms' return home. Use only the robot tools. Act autonomously and ne...
[ { "camera": "head", "url": "https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/videos/s0_l000_arrange_largest_number/head.mp4", "bytes": 1064702 }, { "camera": "left_wrist", "url": "https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main...
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_arrange_largest_number.jsonl.gz
s0_l000_arrange_largest_number_random
gpt-6-astra
arrange_largest_number_random
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
arrange_largest_number
random
Generalization
eval
codex
high
success
1
430.492
{ "take_action_cnt": 674, "action_budget": 1050, "ended_by": "official_end" }
{"start_epoch_s":1792100435,"end_epoch_s":1792100865,"total_wall_s":430.492,"agent_s":276.905,"sim_s(...TRUNCATED)
{ "input": 1565719, "cached_input": 1509504, "output": 5421, "reasoning": 2611 }
{ "calls": 32, "observe": 0, "measure": 9, "run": 23, "step": 0, "assess": 0, "action_status": 0, "errors": 5 }
["I’ll place the digits in descending order on the four pads, check that they’re upright, then r(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
"Arrange the numbers from left to right to form the largest possible number, and place them on the p(...TRUNCATED)
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_arrange_largest_number_random.jsonl.gz
s0_l000_build_tower
gpt-6-astra
build_tower
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
build_tower
Precision
eval
codex
high
fail
0.3
756.462
{ "take_action_cnt": 943, "action_budget": 1050, "ended_by": "agent_end" }
{"start_epoch_s":1792100826,"end_epoch_s":1792101583,"total_wall_s":756.462,"agent_s":445.665,"sim_s(...TRUNCATED)
{ "input": 3440003, "cached_input": 3350016, "output": 10486, "reasoning": 5807 }
{ "calls": 54, "observe": 0, "measure": 22, "run": 31, "step": 0, "assess": 1, "action_status": 0, "errors": 7 }
["I’ll build the two pillar-and-board levels first, then add the brown and green pieces and return(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
Build a tower using the wooden blocks and wooden boards.
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_build_tower.jsonl.gz
s0_l000_classify_objects
gpt-6-astra
classify_objects
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
classify_objects
Long-Horizon
eval
codex
high
success
1
434.515
{ "take_action_cnt": 762, "action_budget": 1100, "ended_by": "official_end" }
{"start_epoch_s":1792100826,"end_epoch_s":1792101261,"total_wall_s":434.515,"agent_s":254.011,"sim_s(...TRUNCATED)
{ "input": 1411239, "cached_input": 1346432, "output": 4868, "reasoning": 1836 }
{ "calls": 30, "observe": 0, "measure": 3, "run": 27, "step": 0, "assess": 0, "action_status": 0, "errors": 7 }
["I’ll sort the three figures, two pens, and two watches into separate baskets, then return both a(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
Sort the objects by category into the three baskets.
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_classify_objects.jsonl.gz
s0_l000_classify_objects_by_language
gpt-6-astra
classify_objects_by_language
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
classify_objects_by_language
Open
eval
codex
high
success
1
575.493
{ "take_action_cnt": 934, "action_budget": 1100, "ended_by": "official_end" }
{"start_epoch_s":1792100907,"end_epoch_s":1792101483,"total_wall_s":575.493,"agent_s":363.615,"sim_s(...TRUNCATED)
{ "input": 2369669, "cached_input": 2265472, "output": 6695, "reasoning": 2844 }
{ "calls": 44, "observe": 0, "measure": 10, "run": 34, "step": 0, "assess": 0, "action_status": 0, "errors": 3 }
["I’ll sort both cars into the left basket, both watches into the middle basket, and both wooden l(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
"Put car objects into the left basket, watch objects into the middle basket, and wooden_toy objects (...TRUNCATED)
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_classify_objects_by_language.jsonl.gz
s0_l000_cover_blocks
gpt-6-astra
cover_blocks
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
cover_blocks
Memory
eval
codex
high
success
1
357.81
{ "take_action_cnt": 499, "action_budget": 800, "ended_by": "official_end" }
{"start_epoch_s":1792101300,"end_epoch_s":1792101658,"total_wall_s":357.81,"agent_s":209.953,"sim_s"(...TRUNCATED)
{ "input": 1293074, "cached_input": 1245312, "output": 3869, "reasoning": 1265 }
{ "calls": 28, "observe": 0, "measure": 10, "run": 18, "step": 0, "assess": 0, "action_status": 0, "errors": 1 }
["I’ll cover left red, middle blue, then right green, and uncover red, green, then blue before ret(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
"Cover the blocks from left to right, remember their colors, then uncover them in the order: red, gr(...TRUNCATED)
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_cover_blocks.jsonl.gz
s0_l000_deposit_coin
gpt-6-astra
deposit_coin
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
deposit_coin
Precision
eval
codex
high
fail
0.2
233.793
{ "take_action_cnt": 245, "action_budget": 300, "ended_by": "agent_end" }
{"start_epoch_s":1792101522,"end_epoch_s":1792101756,"total_wall_s":233.793,"agent_s":177.566,"sim_s(...TRUNCATED)
{ "input": 518978, "cached_input": 476672, "output": 4176, "reasoning": 2770 }
{ "calls": 15, "observe": 0, "measure": 3, "run": 11, "step": 0, "assess": 1, "action_status": 0, "errors": 3 }
["I’ll lift the coin from its holder, align it with the bank’s slot, and return both arms home a(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
Pick up the coin from the holder and insert it precisely into the coin bank.
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_deposit_coin.jsonl.gz
s0_l000_fasten_screws
gpt-6-astra
fasten_screws
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
fasten_screws
Precision
eval
codex
high
success
1
419.054
{ "take_action_cnt": 573, "action_budget": 1900, "ended_by": "official_end" }
{"start_epoch_s":1792101620,"end_epoch_s":1792102039,"total_wall_s":419.054,"agent_s":267.455,"sim_s(...TRUNCATED)
{ "input": 1423839, "cached_input": 1359872, "output": 4944, "reasoning": 2165 }
{ "calls": 31, "observe": 0, "measure": 11, "run": 20, "step": 0, "assess": 0, "action_status": 0, "errors": 2 }
["I’ll match each loose nut to its same-color post, press it down, and return both arms home with (...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
Insert and tighten each screw into the nut of the same color.
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_fasten_screws.jsonl.gz
s0_l000_fill_egg_holder
gpt-6-astra
fill_egg_holder
0
reasoning effort high; codex 0.160.1; pixel coordinates
0
fill_egg_holder
Long-Horizon
eval
codex
high
fail
0.4
400.756
{ "take_action_cnt": 682, "action_budget": 700, "ended_by": "agent_end" }
{"start_epoch_s":1792101696,"end_epoch_s":1792102096,"total_wall_s":400.756,"agent_s":266.623,"sim_s(...TRUNCATED)
{ "input": 1506434, "cached_input": 1420416, "output": 4421, "reasoning": 1678 }
{ "calls": 31, "observe": 0, "measure": 8, "run": 22, "step": 0, "assess": 1, "action_status": 0, "errors": 14 }
["I’ll place all four eggs into separate cups with the right arm, then close the lid with the left(...TRUNCATED)
[{"step":0,"item_id":"","kind":"launch","tool":"","arguments_json":"","text":"\n# This episode\n\nTa(...TRUNCATED)
Place the four eggs from the basket into the egg holder, then close the lid.
"You control the two ARX X5 arms (left and right) in a simulated RoboDojo tabletop episode. Complete(...TRUNCATED)
[{"camera":"head","url":"https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve(...TRUNCATED)
https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim/resolve/main/traces/s0_l000_fill_egg_holder.jsonl.gz
End of preview. Expand in Data Studio

Agent as Policy — Simulation Episodes

Paper · Project page · Code · Real-robot dataset

Formerly agent-as-policy-robodojo (old links redirect). The RoboDojo configurations are now robodojo (was episodes) and robodojo_tasks (was tasks); the RoboDojo release as first published is the tag robodojo-seed0-v1.

Summary

17,397 simulated manipulation episodes in which a language model acts directly as the robot's policy: it sees the cameras and the arm state through tools, measures points in the images, sends motion commands and decides when it is done. There are no trained action heads. Five simulators, one configuration each, plus the results table. One row per episode: the first and last frame, a preview video, the outcome, timing, token and tool counts, every message of the agent and every tool call with its result. Beside the tables: every image the model received, the camera recordings, the model-visible record of each episode (for VLM fine-tuning) and the simulator-side command record (for replay, e.g. to generate step-wise VLA data).

Configurations

config simulator split rows successes models
robodojo (default) RoboDojo (dual-arm ARX X5) evaluated 2,100 1,141 gpt-6-astra
libero LIBERO / LIBERO-PRO / LIBERO-Plus collect 3,300 2,424 gpt-6-astra 3,300
vla_arena VLA-Arena collect 1,780 1,534 gpt-6-astra 1,780
vla_arena VLA-Arena eval 5,995 2,762 gpt-6-astra 1,700, AGP_Qwen3.6 (SFT) 1,700, Qwen3.6-35B-A3B 895, Qwen3.8-Flash-Next 1,700
robotwin RoboTwin 2.0 collect 2,173 1,627 gpt-6-astra 2,173
robotwin RoboTwin 2.0 eval 699 337 AGP_Qwen3.6 (SFT) 300, Qwen3.6-35B-A3B 99, Qwen3.8-Flash-Next 300
robocasa RoboCasa collect 1,350 653 gpt-6-astra 1,350
robodojo_tasks RoboDojo evaluated 42 the official per-task cells
results VLA-Arena, RoboTwin cells 8 one row per cell of the table below
  • collect: trajectories collected with gpt-6-astra (codex, reasoning effort high) to train on — many tasks, about ten initial states each. Some runs used twice the official step limit, some the real-robot-aligned variant (the episode ends when the agent's own completion check is accepted, not at the benchmark's first success); protocol, step_limit_scale, real_end and task_notes say which.
  • eval: the episodes counted in the results table below, under one protocol per benchmark. Only counted episodes are published; interrupted attempts that were run again are not.
  • model is the model that acted; model_settings gives its reasoning or thinking setting, serving and image-coordinate convention.

Results

First-attempt success rate, recomputed from the results table and the episode tables (counted_in names the cell of each counted row).

gpt-6-astra AGP_Qwen3.6 (SFT) Qwen3.6-35B-A3B Qwen3.8-Flash-Next
VLA-Arena 1201/1700 (70.6%) 977/1700 (57.5%) 11/895 (1.2%) partial¹ 573/1700 (33.7%)
RoboTwin 2.0 825/1000 (82.5%) ² 187/300 (62.3%) 6/99 (6.1%) partial¹ 144/300 (48.0%)

¹ Partial baselines, stopped early: VLA-Arena inits 10–14 complete and 45 episodes of init 15; RoboTwin index 013 only. ² A different protocol: seeds 0–9 of the collection without task notes (these rows are in robotwin / collect).

Evaluation protocol (vla_arena_uniform_v1, robotwin_uniform_v1):

  • VLA-Arena: 170 configurations × initial states 10–19 = 1,700 episodes; task notes on; the official step limit; the episode ends at the benchmark's first success. A request that exceeds the remaining step budget runs to the limit and the episode fails there; context cap 131,072 tokens, image cap 256 per request (reaching either is a failure).
  • RoboTwin 2.0: 50 tasks × clean / randomized × seed-ledger indices 013 / 015 / 016 = 300 episodes (seeds checked by the official expert and disjoint from the collection's 0–12); aloha-agilex, 640×480; task notes on; official step limit; ends at the first success; same budget rule; context cap 225,000 tokens, image cap 512. 185 of the 300 SFT episodes ran under earlier, lower caps (131,072 or 258,000 tokens; 256 images) and are counted because none of them reached a cap of the final rule; the ones that ended at an earlier cap were run again. eval_rules gives the caps each row ran with.

Models. gpt-6-astra: codex 0.160.1, reasoning effort high. AGP_Qwen3.6 (SFT): Qwen3.6-35B-A3B fine-tuned on successful gpt-6-astra collection episodes (no reasoning text), served with vLLM, thinking off. Qwen3.6-35B-A3B: the base model, same serving. Qwen3.8-Flash-Next: the NVFP4 release, served with vLLM (fp8 KV cache, no multi-token prediction), thinking on at medium effort with earlier thinking kept in context, and image positions on a 0–1000 scale (image_coords = norm1000); the other models use pixel coordinates. Coordinates in the records are kept exactly as each model read and wrote them.

Files beside the tables

Per simulator, everything an episode has lives in the shards with the same name as its table shard (<sim>/<kind>/<split>-NNNN.*); the row names them.

path what
<sim>/data/<split>-NNNN.parquet the table, one row per episode (one row group per row)
<sim>/views/<split>-NNNN.tar WebDataset: every image the model received, <episode_id>_<NNNN>.png with a .json sidecar (source, tool call, size, sha256); the keys are traces[].image_keys and the rollouts' view_key
<sim>/videos/<split>-NNNN.tar WebDataset: the camera recordings, <episode_id>.<camera>.mp4, one frame per control step or action
<sim>/rollouts/<split>-NNNN.jsonl.gz the model-visible record of each episode (see below)
<sim>/traces/<split>-NNNN.jsonl.gz the agent's event stream (tool results as the harness logged them, images replaced by <image KEY>)
<sim>/replay/<split>-NNNN.jsonl.gz the simulator-side record for replay (see below)
<sim>/robot_profiles/<sha256>.json the robot profile (kinematics, cameras, link hulls) the rows point to
results/cells-0000.parquet the results config
episodes/, tasks/, unstable/, videos/, traces/, views/, documents/, results/seed0/ RoboDojo, see docs/ROBODOJO.md

Sizes over the four simulators: tables 2.63 GiB, views 71.74 GiB, recordings 22.31 GiB, rollouts 0.60 GiB, event streams 0.07 GiB, replay records 0.85 GiB. Column reference and record formats: docs/SIMULATORS.md.

Using the data

from datasets import load_dataset

ds = load_dataset("Agent-as-Policy/agent-as-policy-sim", "robotwin", split="eval")
row = ds[0]
row["model"], row["outcome"], row["counted_in"], row["frame_last"].size
for step in row["traces"]:
    print(step["kind"], step["tool"], step["arguments_json"][:80], step["image_keys"])

cells = load_dataset("Agent-as-Policy/agent-as-policy-sim", "results", split="cells")

VLM fine-tuning (rebuilding the model's context). rollouts/ holds, per episode and in order, the records the agent harness kept of the conversation: session_meta (with base_instructions, the system prompt), every response_item (developer, user and assistant messages; function_call and custom_tool_call with their outputs, matched by call_id; plain-text reasoning where the model produced it) and token_usage_record. Every input_image carries a view_key instead of the image data; the image is <view_key>.png in the row's views_shard. Rebuilding the context means reading the records in order and putting the images back at their keys. The last tool call of an episode can lack an output (the benchmark ended the episode while it ran). Encrypted reasoning and account fields are not included.

VLA data (replay in the simulator). replay/ holds, per episode: an init record (the benchmark's episode key and seed, the backend configuration and switches, the robot profile's sha256 and per-episode calibration notes, the code revision), every request the agent's tools sent to the simulator and its reply (kind = bridge: observe, actions with per-arm motion commands, estimate), and the runtime journal's actions and observations. Building the same episode and sending the requests in order reproduces the trajectory; rendering every control step on the way gives step-wise observation–action pairs.

We replayed episodes of each simulator this way from the published records alone: LIBERO / LIBERO-PRO / LIBERO-Plus 3/3 (largest pose difference 0.000 mm); RoboCasa 3/3 (largest pose difference 0.000 mm); RoboTwin 2.0 3/3 (largest pose difference 0.039 mm); VLA-Arena 3/3 (largest pose difference 0.000 mm). "Reproduced" means every reply matched the recorded one (poses within 1 mm, same motion outcome and step counts) and the benchmark gave the same verdict at the same step. Details per episode: docs/SIMULATORS.md.

RoboDojo

2,100 simulated episodes of the RoboDojo benchmark in which an LLM agent acts directly as the policy of a dual-arm ARX X5 robot: gpt-6-astra, running in codex at reasoning effort high, sees the official RGB cameras and the arms' state through MCP tools, measures points in the images, sends multi-step motion programs and judges its own completion. There are no trained weights. The release is the complete seed 0 of the official evaluation as it was scored: 42 tasks × 50 episodes (the 12 Generalization tasks as 25 standard + 25 randomized layouts) — exactly the episodes behind the official summary below. One row per episode, and the row replays it: every message of the agent, every tool call with its arguments and its result, and a key for every image the model received.

Generalization Precision Long-Horizon Memory Open Average
success rate (%) 54.67 44.00 28.75 91.00 62.25 56.13
score 59.79 50.90 49.48 92.22 65.75 63.63

Per-task results, how seed 0 was assembled, the column reference and the loading examples: docs/ROBODOJO.md, docs/SCHEMA.md, docs/DETAILS.md.

Privacy

Account identifiers, encrypted reasoning, machine paths and host names are removed or replaced by neutral paths (/session, /codex_home, /home/user) in every published text, and every file was scanned for credentials, account values, private paths and host names before upload.

Attribution and licence

The tasks, scenes, assets and evaluators belong to their benchmarks — RoboDojo, LIBERO and its LIBERO-PRO and LIBERO-Plus variants, VLA-Arena, RoboTwin 2.0 and RoboCasa — and their licences continue to apply to everything rendered in them (images, crops, recordings). The messages, tool calls, records and tables were produced by the models named in each row, through codex. Our annotations and tables: CC BY 4.0.

Contributions

Mengzhao Jia (@JillJia) · Yang Lin (@elsannaly) · Xixin Zhang (@XixinZhang)

Citation

@article{jia2026agentaspolicy,
  title   = {Agent as Policy for Robotic Manipulation},
  author  = {Jia, Mengzhao and Lin, Yang and Zhang, Xixin and Zhang, Zhihan and Liu, Xiaobai and Jiang, Meng},
  journal = {arXiv preprint arXiv:2609.12541},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.12541}
}

@dataset{jia2026agent_as_policy_sim,
  title     = {Agent as Policy --- Simulation Episodes},
  author    = {Jia, Mengzhao and Lin, Yang and Zhang, Xixin and Zhang, Zhihan and Liu, Xiaobai and Jiang, Meng},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Agent-as-Policy/agent-as-policy-sim}
}
Downloads last month
4,612

Paper for Agent-as-Policy/agent-as-policy-sim