--- license: cc-by-4.0 pretty_name: RAX robot session logs (Franka, RB-Y1, HumanoidArena) tags: [robotics, llm-agent, claude-code, franka, rainbow-rby1, manipulation, mobile-manipulation, humanoid, simulation] --- # RAX robot session logs - arXiv: https://arxiv.org/abs/2610.22496 - Project page: https://robot-agent-exploration.com/ - Code: https://github.com/robot-agent-exploration/rax - Skill library: https://github.com/robot-agent-exploration/skills-lib Complete logs of Claude Code agents learning manipulation tasks on two real robots and a simulated humanoid through RAX. Each session holds: - the agent's full transcripts, including subagents and tool results; - every robot tool call with its arguments, result and error; - the robot state log and camera recordings; - the operator's instructions; - the code, notes and skills the agent wrote. - **Franka** (2026-10-03 to 2026-10-10): 5 sessions on a Franka arm with a parallel gripper, two exterior ZED stereo cameras and a wrist camera: find a hidden cube, load an air fryer, wipe a table, insert pegs, rebuild a Bandu tower. - **RB-Y1** (2026-10-08 to 2026-10-10): 7 sessions on a Rainbow Robotics RB-Y1 wheeled bimanual manipulator with a head RGB-D camera: move a box and a bucket onto a cabinet, tuck an office chair, stack cups, and a compositional task that chains them. - **HumanoidArena** (2026-10-12 to 2026-10-13): 3 sessions on a simulated Unitree G1 humanoid in HumanoidArena, without expert action demonstrations: score a goal, pick and place a box, and move a hammer to a basket on another desk. Their layout is described in [`humanoid-arena/README.md`](humanoid-arena/README.md). ## Layout ``` README.md this card blur-manifest.jsonl every media file and embedded image changed by face blurring franka//README.md outline and statistics of the session franka//.tar the session directory without camera videos franka//.video.tar camera videos: archive//.mp4 + .json frame times franka/demos/_demo.mp4 edited demo videos (find_the_cube, air_fry_pan, swipe_table, peg_insertion) rby1//... as for franka; .video.tar only for sessions that have recordings rby1/extras/skill-library.zip the RB-Y1 skill library repository (with history) after the last session rby1/extras/skill-library-commit-ids.tsv old -> new commit ids of the rebuilt skill-library histories (see below) rby1/extras/robot-models.zip the vendor URDFs and meshes the sessions used rby1/extras/characterization.zip hardware test results from bring-up rby1/extras/phone-videos/rby1-{1,2,3}.mp4 phone videos filming the compositional session (no audio) humanoid-arena/README.md layout of the HumanoidArena sessions humanoid-arena//.tar one HumanoidArena session (football, pickplace_box, doubledesk) ``` Inside a Franka or RB-Y1 session tar (the HumanoidArena tars are described in `humanoid-arena/README.md`): | Path | Contents | |---|---| | `claude-home/projects/*/.jsonl` | Claude Code transcripts: every message, tool call and tool result (images inline as base64); `*//subagents/` holds subagent transcripts | | `claude-home/history.jsonl` | what the operator typed into the agent CLI, verbatim apart from masking | | `claude-home/settings.json`, `mcp.json` | the agent's permissions, sandbox and MCP server configuration | | `mcp.jsonl` | every robot tool call (`mcp_request`), result (`mcp_result`) and error (`mcp_error`, with type, stage, reason and robot state), epoch seconds | | `archive/robot/.jsonl` | robot state log: Franka at 60 Hz (TCP pose, joints, joint velocities, gripper width), RB-Y1 at 10 Hz (torso, arms, head, end effectors, odometry, wrist forces) | | `agent-workspace/` | AGENTS.md, TASK.md, skills, `memory/` (PROGRESS, RUNS, TIPS), `code/`, `output/`, `skill-library/`, and saved observations in `runtime/frames//` (RGB, depth in uint16 mm, `meta.json` with intrinsics, camera pose and joints) | | `operator/requests/` | requests the agent sent to the operator (resets, help) and the replies | | `configuration.json`, `operator-input.json`, `workspace-manifest.json` | the task, robot, tool contract and session budget | | `agent-lifecycle.json`, `state.json`, `model-usage.jsonl`, `model-calls.json` | agent runs and resumes, token usage per model call | | `frontend.log`, `camera-server.log` | harness and camera service logs | The `.tar` files are uncompressed; extract with `tar -xf`. Transcripts and logs are plain JSONL. ## Sessions ### Franka | Session | Task | Start (UTC) | Agent h | Model calls | Robot tool calls (errors) | Typed instr. | Video | |---|---|---|---|---|---|---|---| | [`20261003T235019Z_find_the_cube_aa6d82`](franka/20261003T235019Z_find_the_cube_aa6d82/README.md) | Find the cube | 2026-10-04 00:07 | 2.7 | 311 | 1767 (14) | 16 | yes | | [`20261004T131834Z_air_fry_pan_dbb6a4`](franka/20261004T131834Z_air_fry_pan_dbb6a4/README.md) | Air fryer | 2026-10-04 13:18 | 2.7 | 365 | 1654 (21) | 5 | yes | | [`20261006T132109Z_swipe_table_9f3e18`](franka/20261006T132109Z_swipe_table_9f3e18/README.md) | Wipe the table | 2026-10-06 13:21 | 3.4 | 441 | 3063 (21) | 53 | yes | | [`20261007T024357Z_peg_insertion_cc65b7`](franka/20261007T024357Z_peg_insertion_cc65b7/README.md) | Peg insertion | 2026-10-07 02:44 | 4.5 | 520 | 5785 (2241) | 47 | yes | | [`20261010T151340Z_bandu_ff0fa2`](franka/20261010T151340Z_bandu_ff0fa2/README.md) | Bandu tower | 2026-10-10 15:13 | 6.1 | 939 | 4511 (23) | 46 | yes | ### RB-Y1 | Session | Task | Start (UTC) | Agent h | Model calls | Robot tool calls (errors) | Typed instr. | Video | |---|---|---|---|---|---|---|---| | [`20261008T022520Z_pnp_box_249b23`](rby1/20261008T022520Z_pnp_box_249b23/README.md) | Box to table (first attempt) | 2026-10-08 02:26 | 0.8 | 89 | 46 (2) | 4 | no | | [`20261008T212446Z_pnp_box_f7b297`](rby1/20261008T212446Z_pnp_box_f7b297/README.md) | Box to table | 2026-10-08 21:25 | 2.5 | 243 | 199 (6) | 17 | no | | [`20261009T000539Z_pnp_box_f75f61`](rby1/20261009T000539Z_pnp_box_f75f61/README.md) | Box to table, then the bucket | 2026-10-09 00:34 | 2.0 | 352 | 1169 (33) | 22 | no | | [`20261009T023700Z_practice_pnp_box_61befa`](rby1/20261009T023700Z_practice_pnp_box_61befa/README.md) | Practice and refactor the box skills | 2026-10-09 02:37 | 1.7 | 311 | 752 (13) | 32 | no | | [`20261009T151644Z_pnp_orange_basket_54c973`](rby1/20261009T151644Z_pnp_orange_basket_54c973/README.md) | Bucket onto the cabinet | 2026-10-09 15:16 | 1.7 | 289 | 450 (18) | 15 | no | | [`20261009T185605Z_push_chair_4ab505`](rby1/20261009T185605Z_push_chair_4ab505/README.md) | Office chair, pick-and-place, cup stacking | 2026-10-09 18:57 | 5.0 | 802 | 2205 (104) | 69 | yes | | [`20261010T011251Z_compositional_manipulation_aa1adf`](rby1/20261010T011251Z_compositional_manipulation_aa1adf/README.md) | Compositional task | 2026-10-10 01:13 | 6.9 | 1335 | 5364 (55) | 60 | yes | **Recordings.** RB-Y1 sessions before 2026-10-09 18:56 UTC (all `pnp_box`, `practice_pnp_box` and `pnp_orange_basket` sessions) have no recordings: no head-camera video and no continuous robot-state log. For those, the robot record is `mcp.jsonl` (every tool call and its result, including the robot state each result reports) and the frames the agent saved. All Franka sessions and the later RB-Y1 sessions have full recordings. **Pruned RB-Y1 sessions.** These 13 RB-Y1 session directories are not included because they hold no agent work on a task: - `20261008T005236Z_motion_check_c02adc`: operator motion check, no agent - `20261008T013827Z_sandbox_audit_989f3c`: harness sandbox test; the agent failed to authenticate - `20261008T013848Z_look_point_wave_2ca762`: never started (configuration only) - `20261008T014715Z_box_side_hand_1a1274`: never started (configuration only) - `20261008T014806Z_isolation_check_4230e6`: harness isolation test (no robot motion) - `20261008T014928Z_isolation_check2_13f35e`: harness isolation test (no robot motion) - `20261008T015244Z_floor_box_hand_65e4fa`: 4-minute stub; disarmed before any arm motion - `20261008T032840Z_health_check_23d66f`: operator health check, no agent - `20261009T000611Z_pnp_box_4a0c2d`: aborted after 7 s, no tool call - `20261009T000745Z_pnp_box_6e2a1a`: 5-minute restart, superseded by pnp_box_f75f61 - `20261009T001027Z_drive_back_917d20`: operator drive, no agent - `20261009T002555Z_controller_check_fae961`: operator controller characterization, no agent - `20261009T232457Z_stack_bowl_and_place_basket_bad3d4`: 1-minute start, superseded by the compositional session ### HumanoidArena | Session | Task | Exploration budget | Formal success | |---|---|---|---| | `humanoid-arena/football` | Football: score a goal | 64 | 59 / 60 | | `humanoid-arena/pickplace_box` | P&PBox: pick and place a box | 128 | 42 / 60 | | `humanoid-arena/doubledesk` | DoubleDesk: move a hammer to a basket on the other desk | 256 | 28 / 60 | ## Privacy and redactions - **Faces are blurred.** Faces were detected with YuNet at full resolution in: - every video: every 2nd frame for the Franka cameras, demo and phone videos, and every frame for the RB-Y1 head camera; - every image, upscaled to at least 1280 px wide; - every base64 image inside transcripts and `mcp.jsonl`. Each detection (score at least 0.65) is pixelated and blurred with padding. In videos the blur also covers 0.5 s either side. Blurred videos are re-encoded as H.264 with the same frame count, so the `.json` frame times still apply. Embedded images are replaced in place, so every JSONL line still parses. `blur-manifest.jsonl` lists every changed file, with the SHA-1 of each original embedded image that was replaced. The detector can miss faces, such as profiles and very small or occluded faces; please report any you find. Some robot parts were blurred by mistake, mostly the RB-Y1 head display in the phone videos. - **Skill-library history.** Three images in the RB-Y1 skill library show people. They are blurred in the working tree and in the git history. The `agent-workspace/skill-library/.git` repositories of the `push_chair` and `compositional_manipulation` sessions and `rby1/extras/skill-library.zip` were rebuilt with the blurred images. Their reflogs were carried over. Commits from the one that first added such an image onward have new ids. Commit ids quoted in the transcripts may refer to the old ids; `rby1/extras/skill-library-commit-ids.tsv` maps old to new. Messages, authors and dates are unchanged. - **Phone videos.** The three phone videos are re-encoded as H.264 with faces blurred, audio removed and all metadata (location, device) stripped. They keep every frame, the average frame rate and the HDR (BT.2020 HLG) colour of the recording. - **Paths.** Lab paths and user names are rewritten. The robot runtime is `/opt/rax/runtime`, the harness checkout `/opt/rax`, the host home `/home/robot`. Other projects that appear in the agent's sandbox deny-list are shown as `other-project-N` / `other-data-N`. In the HumanoidArena transcripts, host storage paths and user and group ids are shown as ``, `` and ``. - **Accounts.** Agent account emails are redacted. The agent CLI's local state is not included: account file, backups, plugin marketplace copies, caches, shell snapshots, job state, task lists and file-edit backups. No API keys or tokens were present. - **Language.** One profane word in the operator's prompts is masked (first letter kept). Everything else is verbatim. ## Notes - `mooncake` is the harness's internal name, kept as recorded (for example, `MOONCAKE_*` environment variables and `harness@mooncake.local` git identities). The public RAX code calls the same things `RAX_*`. - File and directory names use UTC. The agents' notes (`memory/*.md`) and the operator's messages use the lab's local time (UTC-4). - `__pycache__` folders are left out. One agent repository (`code/basket_pnp` in `pnp_orange_basket`) had committed them, so its `git status` lists them as deleted. - `mcp_error` events are kept as the backend returned them. They include refusals by the safety checks (joint limits, self-collision, force and torque guards) and timeouts the agent set itself. Each session README tallies them by tool and reason. - To replay a transcript with Claude Code, place the session at `/opt/rax/runtime/sessions/` and run `CLAUDE_CONFIG_DIR=/claude-home claude --resume` in `/agent-workspace`. The robot tools need the original robot and harness. ## Citation ```bibtex @article{huang2026scaling, title={Scaling Online Exploration for Robot Agents}, author={Huang, Runhan and Zhang, Xiangcheng and Chen, Haoyu and Li, Yitong and Wang, Yen-Jen and Lyu, Bohan and Ye, Haotian and Hong, Chuye and Wang, Yixuan and Su, Chloe and others}, journal={arXiv preprint arXiv:2610.22496}, year={2026} } ```