TacWAM wet-lab tasks — teleoperated xArm6 + BrainCo Revo2 with tactile sensing
Multi-task successor of SingleBicycle/tacwam-pick-up-bottle.
Right arm and right hand only for now; the left side is planned and the layout leaves room for it.
Status (2026-09-04): first delivered batch — smoke_test/, 53 episodes of task S1 cap_to_tray.
Right arm and right hand only. Recorded 2026-09-03/04 in blocks 05–11, success-only selection out of
92 labelled trials (81 % success rate); episodes whose QC gate failed are excluded and stay at the
source. Read Training contract below before writing a loader — three of the twelve action
dimensions are constant in this batch, and whole-episode training is wrong for this data.
On-robot rollouts — rollouts/
Policies trained on this data, run back on the robot, with every layer of the control stack logged: the policy's raw normalized output, the decoded absolute target, the post-guardrail command actually sent, and the measured arm and hand state — plus both camera streams, recorded inside the streaming server so the video is the same bytes the policy was served.
rollouts/smoke_test_20260908/ — 7 runs of S1 cap_to_tray (6 policy, 1 scripted
control). Over 2787 ticks the guardrail never fired, hand clipping changed the request by
0.0 counts, and the arm kept 0.989–1.005 of its commanded travel; the scripted control
grasped and lifted through that same path. The policy's own output closes 23.1 mm above
the taught grasp height. See that folder's README.
Note for loaders: rollouts live at the top level on purpose. load_episode.py globs
smoke_test/* and treats every directory there as an episode, so a rollout folder inside
smoke_test/ would be read as a malformed episode.
Rig state as of 2026-09-03 (arm A, right hand)
| value | how it was measured | |
|---|---|---|
| base frame | arm A base, mm; bench top z = 0; +x forward, +y left, z up | — |
| head camera | base [3, 313, 438] mm, looking +x and 36° down; optical axis meets the bench at [607, 393] | head_cam_solve.py on a 120 s teleop flight: yellow wrist marker in 255 frames vs the arm log, 254 inliers, 2.5 px rms (protocol/extrinsics_head_camera.json) |
| work pose | flange [282, 103, 150], palm down, fingers 20° left of +x | turned +20° about the base for the camera; J1 symmetry measured (identical reach at every angle) |
| teleop box (flange) | [222, 490, −90, 310, 95, 210] | axis-aligned around the props and the rest spot; every corner IK-walked from the work pose, no reconfiguration |
| rest / trial-start spot | flange [290, −60, 210] | hand at the right edge of the frame; both T1 props 100 % visible, robust to ±100 mm camera / ±30 mm prop error (rest_spot.py) |
| T1 source | [425, 85] — the loaded beaker in its ring | back-projected from the pre-flight frame, ±3 cm (protocol/layout_T1_2026-09-03.jpg) |
| T1 target | [430, 250] — inside the white tray (tray spans x 280..690, y 140..470) | same |
| spare beakers | park them to the RIGHT (−y) of the source, never further along the fingers | the open fingers reach 236 mm past the flange and rake anything behind the grasped object |
Why a rest spot. At the work pose the palm sits on the camera's line of sight to the source beaker
(0 % visible), so every trial's first frames showed the arm. Each trial therefore starts and ends at the
rest spot: retreat right and up until the box ceiling stops you, then release the clutch. The first
trial of a block is a short positioning move (label discard).
Every trial is kept. trial.label ∈ {success, fail, discard}; a fail carries failure_reason ∈ {grasp, drop, place, collision, tracking, operator, other}; ledger.csv lists all of them with the
plan's condition. Delivery is a selection, never a deletion.
Per-trajectory video (viz/<episode>.mp4, example frame protocol/viz_format_example.jpg): head
camera | wrist camera | tactile pressure field on the 15 glove pads laid out as a hand | per-finger
contact-force traces.
Rig
| arm | UFACTORY xArm6 (arm A, table-mounted; bench top is base z = 0) |
| hand | BrainCo Revo2 right hand, 6 motors, wearing a Tujian EI-G012 tactile glove (15 pads, 880 taxels) + 5 fingertip pads of its own |
| teleop | UDEX gloves + PICO wrist tracker → clutch → Cartesian teleop at 30 Hz |
| cameras | static head RGB-D (Orbbec Gemini 345Lg, not egocentric: base [3, 313, 438] mm, looking forward and 36° down — solved 2026-09-03), wrist RGB (RealSense D405 on the flange) |
| work pose | palm-down, turned +20° about the base for the camera; flange [282, 103, 150]; rest spot [290, −60, 210]; box [222, 490, −90, 310, 95, 210] |
Tasks
| id | instruction (EN) | 指令 (ZH) | conditions | success |
|---|---|---|---|---|
T1 cup_transport → tasks/T1_cup_transport/ (layout images, coordinates, choreography) |
transport the loaded cup to the tray | 搬运装载的杯子到托盘 | L/M/H bead load | cup upright in the tray, hand clear, nothing knocked over |
T2 tube_to_rack |
move the tube from rack A to rack B | 把试管从 A 架移到 B 架 | A→B / B→A | tube standing in the destination rack |
T3 compliance_sorting |
sort the block into the bin matching its compliance | 按软硬把方块放进对应的盒子 | soft / rigid | block in the matching bin, tray unmoved |
T4 tool_bench_cleanup |
wipe the debris out of the marked zone with the scraper | 用刮板把碎屑清出标记区 | 3 lanes / scattered | all pieces outside the zone, tool released upright |
T5 dry_media_transfer |
pour the media into the bowl and return the cup | 把颗粒倒进碗里并放回杯子 | beads / rice | most media in the bowl, cup upright in the finish area |
S1 cap_to_tray (smoke test) |
pick the cap and place it in the white container | 拿起盖子放进白色容器 | cap on its stand | cap inside the container, hand clear |
Every trial starts and ends at a rest spot where the hand is out of the head camera's line of sight to the props, so the first frames show the scene. Prop positions, allowed variation, motion script and failure taxonomy per task: docs/COLLECTION_PROTOCOL.md in the source repository.
Layout
README.md
tasks/<task>/ per-task page: layout images (camera + top view), coordinates, choreography, live yaml
protocol/ COLLECTION_PROTOCOL.md (per-task cards), SMOKE_TEST_RUNCARD.md (bilingual session
card), the live task yamls, work_pose.json, the head-camera solve, a viz example
smoke_test/ 53 episodes, Tujian-G3 shaped — one directory per episode:
episode_30hz.h5 every stream on one 30 Hz grid + valid/ masks
rgb_head.mp4/.csv STATIC table camera (not egocentric)
wrist_right.mp4/.csv D405 on the flange
depth_head.mkv FFV1 lossless GRAY16, mm
right_hand_data.npz Tujian glove tactile, 880 taxels over 15 pads
robot/ arm.jsonl, arm_cmd.jsonl, hand.jsonl, pico.jsonl, manifest.json
task_info.json, qc_gate.json, head_param.json, kalibr_parameters.yaml
plus dataset.csv, metadata.jsonl, delivery_*.csv, ledger.csv, robot_model/,
safety_deployment.json, environment.yml
ledger.csv lists every trial ever recorded for this task — success, fail and discard — with its
block, condition, failure reason and whether it was delivered. The delivered set is a selection, not
the whole of what exists.
Labels
Delivered task names keep the bottle-cap format: name in Chinese (搬运杯子 · 试管插架 · 软硬分拣 · 清扫台面 · 倾倒颗粒物),
name_en = the pack id (cup_nest, tube_to_rack, tray_pack, bench_wipe, dry_pour).
Every trial is kept. trial.label ∈ {success, fail, discard} plus a failure_reason from a fixed
vocabulary (grasp, drop, place, collision, tracking, operator, other). Failures ship
with the same files as successes; consumers filter on the label.
Training contract — post-training a Cosmos3 policy from a mid-train checkpoint
Everything in this section is measured on the delivered smoke_test/ batch (53 episodes,
14 356 action frames inside the clutch window), not assumed. Read it before writing a loader.
0. Verify all of this yourself — do not take our numbers on trust
Every figure in this section is something we measured on our copy of the batch. Re-run it on what you actually downloaded and trust your own output over ours. If your numbers differ from the tables below, your numbers are the ones that matter, and please tell us.
load_episode.py in the repo root is the reference loader. It applies every rule in this section —
the clutch slice (§3), the constant-dimension guard (§1), the whole-block video indices (§7) — and
refuses rather than returning a silently wrong array. Use it, or read it and copy the rules; it is
~90 lines with no dependencies beyond h5py and numpy. Running it prints a summary you can compare
against the tables here:
$ python load_episode.py smoke_test
53 episodes, 14354 task frames (478 s at 30 Hz), action 12-D
constant dims: [3, 4, 5] (scale forced to 1.0; they normalize to exactly 0)
normalized: NaN 0, |z| max 4.57, |z| max over the constant dims 0.00e+00
# reproduces every statistic in this section, ~1 min over the 53 episodes
import glob, warnings, h5py, numpy as np
ACT, STEPS, rev, tot = [], [], 0, 0
for h in sorted(glob.glob("smoke_test/*/episode_30hz.h5")):
with h5py.File(h) as f:
keep = (f["valid/action_arm_right"][:] & f["action/right/clutch"][:]).astype(bool)
aa = f["action/right/arm_target_aa"][:][keep]
if len(aa) < 10:
continue
ACT.append(aa)
d = np.diff(aa[:, :3], axis=0); STEPS.append(np.linalg.norm(d, axis=1))
u, v = d[:-1], d[1:]; n = np.linalg.norm(u, axis=1) * np.linalg.norm(v, axis=1)
ok = n > 1e-6
rev += int(((u[ok] * v[ok]).sum(1) / n[ok] < 0).sum()); tot += int(ok.sum())
A = np.concatenate(ACT); S = np.concatenate(STEPS)
print("episodes", len(ACT), " action frames", len(A))
print("distinct values per dim:", [len(np.unique(A[:, i])) for i in range(6)]) # expect [.., 1, 1, 1]
for name, X in (("float32", A.astype(np.float32)), ("float64", A.astype(np.float64))):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
z = (X - X.mean(0)) / X.std(0)
zr = z[:, 3:] # the three orientation dims
fin = np.isfinite(zr)
print(f"{name}: rot std={X.std(0)[3:]} NaN={np.isnan(z).sum()} "
f"|z|max over rot dims={np.abs(zr[fin]).max() if fin.any() else float('nan'):.3f}")
print(f"step mm: mean {S.mean():.2f} p50 {np.percentile(S,50):.2f} p95 {np.percentile(S,95):.2f}")
print(f"direction reversals {100*rev/max(tot,1):.1f}%")
What we saw, for comparison:
episodes 53 action frames 14356
distinct values per dim: [9564, 10144, 2989, 1, 1, 1]
float32: rot std=[8.41e-03 1.21e-06 1.54e-02] NaN=0 |z|max over rot dims=1.000
float64: rot std=[0. 0. 0.] NaN=43068 |z|max over rot dims=nan
step mm: mean 3.77 p50 2.83 p95 10.86
direction reversals 7.6%
The 1, 1, 1 on the right of the distinct-value line, and the float32 row showing no NaN with
the orientation dims still swinging a full ±1 σ, are the whole of §1 in two lines of output.
1. ⛔ Three of the twelve action dimensions are CONSTANT — and naive normalization turns them into noise
| action dim | mean | std (float64) | min | max | distinct values in 14 356 frames |
|---|---|---|---|---|---|
| x (mm) | 348.99 | 37.83 | 285.00 | 468.01 | many |
| y (mm) | 52.21 | 77.21 | −90.00 | 261.51 | many |
| z (mm) | 216.42 | 33.93 | 120.97 | 338.32 | many |
| rx (deg) | 127.286 | 0 | 127.286 | 127.286 | 1 |
| ry (deg) | 0.009 | 0 | 0.009 | 0.009 | 1 |
| rz (deg) | 127.273 | 0 | 127.273 | 127.273 | 1 |
Wrist-orientation following is deliberately disabled for this task (orientation.enabled: false):
the tool attitude is frozen the moment the operator engages the clutch and never changes, so each of
those three channels holds one bit-identical float32 for every frame of every episode.
The failure is not a divide-by-zero, and that is what makes it dangerous. Measured on the delivered batch:
| how you compute it | result |
|---|---|
| float64 accumulation | std = 0 exactly → 0/0 → 43 068 NaN. Loud, you will see it at step 1. |
| float32 accumulation (the h5 dtype, and the default) | the mean carries rounding error, so std comes out 8.4e−3 / 1.2e−6 / 1.5e−2 and (x − mean) / std lands in ±1.0 — no NaN, no Inf, no warning |
In the float32 path those three channels arrive at the model as full-scale ±1 σ signal that is
pure floating-point rounding noise. Three of the six arm-action dimensions — half of the arm's
action loss — then chase a target no model can predict, and an eps guard does not help, because
std is already 10³–10⁴ times larger than any sensible epsilon.
Do this instead: detect constant channels physically, not numerically.
CONST_TOL_DEG = 0.1 # far below any real wrist motion
sd = act.std(0)
const = sd < CONST_TOL_DEG # or: np.array([len(np.unique(act[:, i])) == 1 ...])
scale = np.where(const, 1.0, sd) # never divide a constant channel by its own noise
z = (act - act.mean(0)) / scale # constant dims become exactly 0
At inference the denormalization returns the frozen attitude, which is the correct command for this task. The physical cost of getting it wrong is negligible (±0.015° of rotation); the training cost is not.
What this actually does to a training run, if you leave it unguarded:
- Nothing visibly breaks. The run completes and the metrics look ordinary. There is no NaN, no warning, nothing in the logs to notice.
- Half of the arm's action loss becomes unlearnable. Three of the six arm dimensions are chasing floating-point rounding noise. Gradients are spent on a target no model can predict, the xyz dimensions learn more slowly for it, and the action loss settles on a floor it cannot go below.
- The robot is unaffected. Denormalized, those channels move ±0.015°, which is nothing. This is a training-efficiency problem, not a safety one — the arm still holds palm-down.
- It teaches "the wrist never turns." For this task that is correct. But T2 (side-grasp, ~90° twist) and T5 (pouring) need wrist rotation, and a checkpoint post-trained only on this batch will have to be retrained on orientation-varying data before it can do them.
Context worth having: the tacwam-pick-up-bottle corpus spans only 1.72° of attitude, so it is
near-frozen too. A mid-train checkpoint built on it has never seen meaningful wrist rotation either —
this batch's exact zero is consistent with that history, not a regression from it.
2. Axis-angle hemisphere — consistent, and it matches the bottle-cap corpus
|aa| = 180.00° on every frame: this attitude sits exactly on the ±180° singularity shell, which is
where a rotation vector has two equal spellings. It does not flip here — of 14 356 frames, 0
have a negative dot product against the batch mean [127.29, 0.01, 127.27]. That reference is the
same one the tacwam-pick-up-bottle corpus uses, so the two are directly mixable.
The packer canonicalizes arm_target_aa at write time. Take your hemisphere reference from the
checkpoint's declared action mean, not from a constant, if you mix in other tasks.
3. ⛔ Slice on the clutch window — do not train on whole episodes
Keep only frames where valid/action_arm_right == 1 AND action/right/clutch == 1.
One block is one continuous capture, hard-linked into every trial; the per-trial split is a window
clamp at align time, so each episode deliberately keeps the lead-in and the operator's own hand
resetting the prop. Measured on this batch: the clutch engages 5.4–16.0 s after the episode opens,
and valid/action_arm_right covers only 30–47 % of the grid (75–83 % of the engaged window). Roughly
half of every episode is outside the action, and under a commanded-state contract a frame with no
command is not a neutral sample — it teaches the policy to wait.
Why the conjunction and not either alone: a re-engage within 3 s counts as a regrip rather than a new
trial, so a fast reset can land inside trial.engaged_host_ns … released_host_ns; and the mask alone
leaks 1–3 zero-order-hold frames at each release boundary (7 of 14 episodes checked). The conjunction
is clean in both cases.
Where the frames go. Measured across the 53 episodes, in seconds:
| segment | mean | p50 | p90 | max | share |
|---|---|---|---|---|---|
| whole episode | 21.8 | 21.7 | 26.0 | 28.6 | 100 % |
| lead-in (episode opens → clutch engages) | 7.3 | 8.1 | 10.1 | 13.3 | 33.5 % |
| task (engage → release) | 11.5 | 11.2 | 13.5 | 19.3 | 52.7 % |
| tail (release → episode closes) | 3.0 | 3.0 | 3.0 | 3.0 | 13.8 % |
The tail is a fixed 3.0 s in every episode: that is the collection tool's --trial-gap, the interval
that has to elapse before a re-engage counts as a new trial rather than a regrip. The lead-in is the
operator swapping the cap back to its start mark and getting clear of the frame. Neither is action.
Inside the task window a further ~20 % of grid frames carry no fresh command, because the arm command
loop runs at a median 25.4 Hz against the 30 Hz grid (p90 inter-command gap 54 ms vs the 33 ms
cell). Those frames are marked invalid rather than filled, which is why valid/action_arm_right
covers 75–83 % of the engaged window and not 100 %. If you would rather hold the last command across
them than drop them, that is a defensible choice under a commanded-state contract — but it is your
choice to make, and the mask is there so you can make it.
import h5py, numpy as np
with h5py.File("smoke_test/<uuid>/episode_30hz.h5") as f:
keep = (f["valid/action_arm_right"][:] & f["action/right/clutch"][:]).astype(bool)
# The HAND command has its own validity and can be invalid on a frame where the arm's
# is fine. Measured on this batch it is 2 frames in 14 356 -- but left as NaN those two
# poison mean() and std() for all six hand columns, and every normalized value in the
# batch comes back NaN. Hold the last commanded value instead, which is what the
# commanded-state contract says the row is.
hand = f["action/right/hand_target"][:]
vh = f["valid/action_hand_right"][:].astype(bool) & ~np.isnan(hand).any(1)
idx = np.maximum.accumulate(np.where(vh, np.arange(len(vh)), -1))
hand = hand[np.clip(idx, 0, None)]
keep &= idx >= 0 # no command yet -> drop
act = np.concatenate([f["action/right/arm_target_aa"][:][keep], # xyz mm + aa deg
hand[keep]], 1) # 6 motors, 0..1000
assert not np.isnan(act).any()
4. Action and state layout — unchanged, and locked
- Action (12-D):
action/right/arm_target_aa(xyz mm + axis-angle deg) ++action/right/hand_target(6 motors, 0..1000). - State row: the same 12 slots, commanded — the last arm target and last hand target, not
measured proprioception. The measured streams (
obs/robot/right/tcp_pos,tcp_quat,joint_pos,hand_pos) are delivered for analysis and QC, and are not the model's input. - Do NOT read
action/right/arm_target_rot6. That field is column-major while the inference config declares row-major, and a loader that prefers the field when present will silently disagree. Treat it as absent.
Widening the layout is what broke the 2026-09-01 checkpoint: a 21-wide state row and 15-wide actions made the pretrained action projections relearn from scratch, and 3 000 fine-tune steps produced a noise-dominated head that aborted all seven live trials. New task = new data + new prompt, same layout.
5. Hand channel notes
| motor | mean | std | max |
|---|---|---|---|
| thumb_flex | 153.1 | 105.9 | 618 |
| thumb_aux | 127.2 | 173.4 | 714 |
| index | 253.4 | 158.3 | 941 |
| middle | 144.1 | 180.3 | 787 |
| ring | 257.3 | 294.0 | 1000 |
| pinky | 251.2 | 369.6 | 1000 |
thumb_aux is the opposition joint. The glove does not sense thumb opposition — the vendor SDK
synthesizes that channel from flexion (ch20 = 4.98 − 0.600 × ch2, r = −0.999997), so thumb_aux here
is a scaled copy of curl, not an independent measurement. Ring and pinky saturate at 1000.
6. Smoothness reference for your offline ship gate
Measured on the demonstrations themselves, inside the clutch window, at 30 Hz:
| row-to-row step |Δxyz| | mean 3.77 mm, median 2.83, p95 10.86, max 31.88 |
| direction reversals | 7.6 % |
Before any robot time, run your offline rollout on at least two training episodes and compare. A checkpoint whose predicted step is ~2× the demonstrations' or whose reversal rate is 40–50 % is noise-dominated; that exact signature cost seven live trials on 2026-09-01. This gate takes five minutes on one GPU.
7. Cameras are NOT egocentric
rgb_head is a static table camera (Orbbec Gemini 345Lg on a tripod, solved at base
[3, 313, 438] mm looking forward and 36° down), despite the name. wrist_right rides the robot
flange. If your mid-train checkpoint was pretrained on egocentric video, condition on this — the
h5 carries the mounts in its camera_mounts_json root attr.
8. Language instruction
task_info.json carries name = 盖子放入容器 (Chinese, the delivered convention) and
name_en = cap_to_tray. The natural-language form is "pick the cap and place it in the white
container" / "拿起盖子放进白色容器". Pick one string per task and keep it fixed — the checkpoint
learns exactly that string.
9. The other risks, ranked by how much they cost you
- Slicing is the big one. About half of every episode carries no action command. Train on whole episodes and the policy learns to wait — far more damaging than the orientation noise above. See §3.
- Success-only delivery. No recovery behaviour is in this batch. The 14 failed trials are kept at
the source with their
failure_reason; ask and we will ship them. - The batch is small. 53 episodes, 14 356 action frames, 8.0 minutes of actual manipulation, averaging 9.0 s per episode. That is the same order as the 47-episode bottle-cap run that post-trained successfully in 3 000 steps — enough to adapt a mid-train checkpoint, not enough to change it substantially.
中文摘要（训练侧）
这一批 53 条轨迹用于 Cosmos3 的后训练，有四件事必须先知道，而且请自己跑 §0 的脚本复核，不要直接采信我们的数字。
仓库根目录的 load_episode.py 是参考读取实现，下面四条它都已经处理好；直接用它，或者读一遍把规则抄进你自己的 loader。它不到 90 行，只依赖 h5py 和 numpy，遇到读不出的情况会直接报错而不是返回一个悄悄错掉的数组。
12 维动作里 rx / ry / rz 是常量（127.286 / 0.009 / 127.273，逐位相同），因为这个任务关掉了手腕跟随，姿态在合离合那一刻冻结。用 float64 求 std 得 0，
0/0产生 43068 个 NaN；但用 float32（h5 原始类型，也是默认路径）均值有舍入误差，std 算出 8.4e−3 / 1.2e−6 / 1.5e−2，(x−mean)/std落在 ±1.0，不报错、不告警。后果：训练照常跑完、指标正常，但手臂 6 维动作里有 3 维在拟合浮点噪声，一半的手臂损失不可学习，位置三维收敛变慢。加eps没用（std 比 eps 大三四个数量级）。正确做法是按物理量级判定常量通道：std < 0.1°即视为常量，scale 取 1.0。对机器人本身无害（反归一化后仅 ±0.015°）。必须按
valid/action_arm_right == 1且action/right/clutch == 1切片，不要拿整段 episode。每段里约一半的帧没有动作命令，喂进去会教模型"等待"。这条比第 1 条更严重。实测（单位秒）：整段平均 21.8，其中前导 7.3（操作者把瓶盖放回起始标记并退出画面）、任务 11.5、尾巴固定 3.0（采集工具的--trial-gap，判定"重新握持"还是"新一条"的间隔）。任务段内还有约 20% 的格点没有新命令，因为手臂命令环实测中位 25.4 Hz 而网格是 30 Hz，这些帧被标为无效而不是填充——要不要用零阶保持补上，由你决定，mask 就是为了让你能做这个决定。另外
hand_target有自己的有效位，可能在手臂有效的帧上无效。这批只有 2 帧（共 14356 帧），但只要留着 NaN，6 个手部维度的mean/std全部变成 NaN，整批归一化结果静默全废。按"最后一次命令"保持住即可，§3 的代码片段已经这么写了。只交付成功轨迹，模型学不到失败恢复；14 条失败留在源端，需要可以补发。
数据量偏小：14356 动作帧、总计 8.0 分钟真实操作。够适配一个 mid-train ckpt，不足以做大幅改动。
补充：瓶盖语料的姿态跨度也只有 1.72°，基本同样冻结，所以第 1 条不是相对该 ckpt 的退化。T2 侧抓和 T5 倾倒需要手腕旋转，届时必须用带姿态变化的数据重训。
- Downloads last month
- 27