SDK Reference
The Python semantic API for Wuji Hand 2. Unified pattern: hand.{resource}().{action}(). All operations target the whole hand (20 joints).
This page focuses on the Wuji Hand 2–specific semantic API. The SDK builds on the shared Wuji SDK (wuji_sdk, pip install wuji-sdk)—for general installation, device connection, and data subscription, see the Wuji SDK docs. For source, see the wuji-sdk repository.
Connection
Network prerequisite: Wuji Hand 2 uses a static IP, not DHCP. Each device ships with a fixed address by handedness — left hand 192.168.1.110, right hand 192.168.1.111, gateway 192.168.1.1, subnet mask 255.255.255.0. Before the first connection, set your host NIC to the same subnet (for example 192.168.1.100), then connect by address. The factory IP is changeable later — see ip (SET/GET).
Connect via SdkManager (Recommended)
from wuji_sdk import SdkManager, Handedness
manager = SdkManager.instance()
# Auto-discover and connect to the first Wuji Hand 2 on the LAN
hand = manager.auto_connect(device_name="wuji_hand_2")
# Connect explicitly by serial number or address
hand = manager.connect(sn=<device_sn>, device_name="wuji_hand_2")
hand = manager.connect(address="192.168.3.110:50001", device_name="wuji_hand_2")
# Or connect by handedness directly, without a serial number
hand = manager.connect(handedness=Handedness.Right, device_name="wuji_hand_2")The device_name="wuji_hand_2" overloads of auto_connect and connect return a WujiHand2 instance directly (with type hints). handedness is mutually exclusive with sn and address. If multiple Hand 2 units of the same handedness are on the network, AmbiguousHandedness is raised—specify sn instead.
ConnectOptions
from wuji_sdk import SdkManager, ConnectOptions
opts = ConnectOptions(
timeout_ms=1000,
retry_count=3,
enable_bridge=True, # Default True: lets multiple clients (Wuji Studio + recording scripts + your app) connect at once
)
hand = manager.connect(sn=<device_sn>, device_name="wuji_hand_2", options=opts)Set enable_bridge=False for exclusive single-client mode.
Instance Attributes
| Attribute | Type | Description |
|---|---|---|
serial_number | str | Device serial number |
device_name | str | Name given at connect (default "wuji_hand_2") |
info | Optional[DeviceInfo] | Device info: serial_number, firmware_version |
is_connected | bool | Connection state |
hand.hw_version().get() returns the factory hardware version HwVersion(major, minor, patch). 0.0.0 means it wasn't written at the factory.
Hand-Level Resources
This section covers the whole-hand API (20 joints). Single-joint reads and actions go through the JointHandle returned by hand.joint(k) / hand.joints() (see Joint Traversal). The device runs in MIT control mode by default — the control mode isn't set from the Python API.
handedness (GET)
Get the handedness result (left / right).
side = hand.handedness().get() # → "left" or "right"online_joints_count (GET)
Get the online joint count (0–20).
n = hand.online_joints_count().get() # → int, 0–20joint_diagnostics (SUB)
Joint diagnostics stream: subscribe to per-joint status word, current, bus voltage, temperature, and error code. Frames are variable-length and contain only online joints — identify each entry by its nid. This stream and joint_states ride the same device stream, so lowering the push rate of one lowers both — see joint_states (SUB).
sub = hand.joint_diagnostics().subscribe() # → Subscription[JointDiagnosticsFrame]
frame = await sub.recv_async()
for j in frame.joints:
print(j.nid, j.vbus_v_fb, j.mcu_temp_c_fb, f"0x{j.error_code_current:04X}")
sub.close()Decode error_code_current with the static method WujiHand2.describe_error(code). The returned object carries the error name, description, severity (Warning / DeferredStop / ImmediateStop / Fatal), and clear_policy (AutoClear / ManualClear / NonClearable). Unknown codes return None. The code field of the returned object is an integer, the same value you passed in. Printing a diagnostics entry object shows error codes in hex.
For joint angles, use position from the joint_states stream — the diagnostics stream doesn't carry joint angles.
comm_diag (GET, 1 Hz)
Communication diagnostics: returns the throughput / error rate of the 5 fingers, once per second.
diag = hand.comm_diag().get() # → HandCommunicationDiagnostics
for finger in diag.fingers: # throughput / error rate for each of the 5 fingers
print(finger.tx_kbps, finger.rx_kbps, finger.error_per_sec)Use it to troubleshoot communication between the joints and the control board.
effort_limit (SET/GET)
Effort limit: read / write the per-joint torque cap (A).
hand.effort_limit().set(1.5) # set all joints to 1.5 A
limits = hand.effort_limit().get() # → list[Optional[float]]Writes reject NaN, Inf, negative values, and values above the device ceiling — whole-hand and per-joint writes validate the same way. A rejected write fails with an error and leaves the current value unchanged. The firmware sets the ceiling. For common errors and remedies, see Wuji SDK troubleshooting.
mit_params (SET/GET)
MIT impedance parameters: read / write per-joint kp / kd.
hand.mit_params().set((1.0, 0.05)) # same kp / kd for all joints
hand.mit_params().set([(1.0, 0.05)] * 20) # list of 20, per joint
mp = hand.mit_params().get() # → list[Optional[MitParam]]; offline joints are NoneWrites reject NaN, Inf, and negative values.
clear_fault (EXEC)
Clear faults: a direct action. Without arguments it acts on the whole hand.
hand.clear_fault() # clear faults on all joints
hand.clear_fault(joints=mask) # optional: 0/1 mask of length 20 — only joints with a 1 are clearedUser Origin
Calibrate the current physical position as the joint-side zero. If a joint is moving, it takes effect at the next IDLE.
hand.set_origin() # whole hand: refresh the user origin of all joints
hand.clear_origin() # whole hand: clear the user origin of all joints
hand.set_origin(joints=mask) # optional: 0/1 mask of length 20 — only joints with a 1 are affectedEnable / Disable / Emergency Stop
hand.enable() # enable all joints
hand.disable() # disable all joints
hand.enable(joints=mask) # optional: 0/1 mask of length 20 — acts on joints with a 1
hand.emergency_stop() # emergency stop: whole-hand action, no maskjoint_states (SUB)
Joint state stream: subscribe to the real-time state frames of the 20 joints, at a native output rate of 1000 Hz. Lower the push rate through the subscription handle: sub.set_rate(hz) returns the rate in effect, and 0 restores the native rate — see Wuji SDK data subscription for full usage. This stream and joint_diagnostics ride the same device stream, so lowering the rate of one lowers both.
Push-rate tuning applies to the joint_states, joint_diagnostics, fingertip tactile, and imu streams. Upgrade the firmware and the Wuji SDK together to versions that support it — see the release notes for the version requirements.
sub = hand.joint_states().subscribe() # → Subscription[JointStateFrame]
frame = sub.recv() # synchronous, non-blocking; None = no data yet
frame = await sub.recv_async() # asynchronous wait
print(frame.header.seq, frame.header.timestamp_us)
for j in frame.joints:
print(j.nid, j.position, j.velocity, j.effort)
sub.close()Callback mode:
def on_state(frame):
print(frame.header.seq, [j.position for j in frame.joints[:4]])
cb = hand.joint_states().subscribe_with_callback(on_state)
# ...
cb.close()position is the joint-side angle (rad) and velocity the joint-side angular velocity (rad/s)—this is the only recommended path for joint angle and speed. Frames are variable-length and contain only online joints — identify each entry by its nid. The joint_diagnostics stream doesn't carry joint angles.
joint_command (PUB)
Joint command: publish positions / velocities / torque feedforward for the 20 joints. Each send takes exactly 20 JointCommand entries, one position / velocity / effort per joint.
from wuji_sdk import JointCommand
pub = hand.joint_command().publish() # → JointCommandPublisher
pub.send([JointCommand(position=p, velocity=0.0, effort=0.0) for p in positions])
pub.close()To turn human hand keypoints into these 20-joint commands, use Wuji SDK Hand Retargeting.
Fingertip Tactile Streams (SUB)
Fingertip tactile requires Beta 2 hardware equipped with the fingertip tactile sensors, plus firmware v2.1.0 and SDK v2026.7.21 or later, upgraded together.
Self-describing per-finger sensor streams: first fetch the finger's FingertipSensorInfo metadata (its format JSON describes the data-frame layout), then subscribe to that finger's FingertipSensorData stream and decode per the format. The thumb has 40 sensing points, the other fingers 34. The native output rate is 100 Hz, adjustable down through the subscription handle.
Starting with firmware v2.4.0, per-point force arrives normalized against the sensor's full scale: 0 to 1 on the normal axis and -1 to 1 on the two tangential axes, with readings past full scale clamped to the limit. Firmware before v2.4.0 reports per-point force in newtons. The three-axis resultant force stays in newtons and the temperature in Celsius, unaffected by this change.
Read scale and unit from info.format when decoding instead of hard-coding either convention. A firmware upgrade changes the info digest, so the digest check in the example rebuilds the decoder and the new range takes effect. Normalization needs Wuji SDK v2026.8.17 or later.
import json
info = hand.get_fingertip_info(0) # 0=thumb … 4=pinky → FingertipSensorInfo
fmt = json.loads(info.format) # the format fully describes the frame layout — never hardcode it
sub = hand.fingertip_thumb_data().subscribe() # one resource per finger: fingertip_{thumb,index,middle,ring,pinky}_data
frame = await sub.recv_async() # → FingertipSensorData
if frame.info_digest != info.digest: # a digest mismatch means the info changed
info = hand.get_fingertip_info(0) # re-fetch and rebuild the decoder
# decode frame.data per fmt's point_fields / aggregate_fields
sub.close()For a complete consumer-side reference (including decoder construction), see the 3.fingertip_typed.py example under examples/python/wuji_hand_2/.
Fingertip Tactile Calibration (EXEC) and Status (GET)
Zero-baseline recalibration: one call recalibrates all 5 fingertips to a fresh zero baseline.
hand.tactile_calibrate() # whole-hand action; fails fast on the first offline fingerKeep all sensor surfaces unloaded (no contact force) during the call — otherwise the calibrated zero baseline is wrong.
A successful call means the calibration command reached every finger, not that calibration finished. Confirm with the status query:
from wuji_sdk import TactileState
status = hand.tactile_status("thumb") # "thumb" / "index" / "middle" / "ring" / "pinky"
print(status.model) # TactileType.Thumb (40 points) or TactileType.Standard (34 points)
print(status.state) # TactileState.Ready / TactileState.Calibratingtactile_status is a blocking query (up to about 1 s). Calibration is complete once polling reads state == TactileState.Ready.
Flash Log Export (Diagnostics)
# This call blocks until the export finishes.
result = hand.export_flash_logs("./logs")
print(result["path"]) # str: absolute path to the exported JSONL file
print(result["frames"]) # int: number of exported log recordsThe SDK writes one flash_<serial>_<YYYY-MM-DD_HHMMSS>.jsonl file to out_dir. Omit out_dir to use ~/.wuji/logs.
The SDK writes to a temporary file and makes the JSONL file available only after the export finishes. If the export fails, the SDK removes the temporary file, so the output directory contains no partial export.
When multiple sessions export logs from the same device, the SDK retries automatically. If the export still can't finish, the call reports that the device is busy. The SDK removes the temporary files and doesn't keep an unfinished log.
Flash log export requires compatible device firmware and Wuji SDK versions. If the versions are incompatible, the export fails without creating log files. Update the device firmware and the Wuji SDK together, then retry.
ip (SET/GET)
Read and write the device's static IP address. set writes the new IP to the device's flash. The firmware doesn't hot-swap the running Ethernet stack, so the new IP takes effect only after the next reboot. Between set and reboot, get still returns the current IP.
hand.ip().get() # read the current IP, e.g. "192.168.3.110"
hand.ip().set("192.168.2.111") # write to flash, takes effect after rebootFull round-trip to change the IP: set → reboot → reconnect at the new IP → get to confirm.
hand.ip().set("192.168.2.111")
hand.reboot()
manager.disconnect_all()
# wait for the device to reboot and Ethernet to come up, about 8 seconds
hand = manager.connect(address="192.168.2.111:50001", device_name="wuji_hand_2")
hand.ip().get() # → "192.168.2.111"The new IP takes effect only after a reboot. set writes to flash without changing the live connection. Before the reboot, get still returns the current IP.
reboot (EXEC)
Reboot the device. The device disconnects and needs a fresh connection. Pair it with ip().set() to apply a new IP written to flash.
hand.reboot()imu (SUB)
Onboard IMU stream: pushes acceleration and angular velocity at a native output rate of 100 Hz, adjustable down through the subscription handle. The device does no onboard orientation fusion, so the orientation field is unavailable and orientation_covariance[0] = -1 marks it invalid per the ROS sensor_msgs/Imu convention. Data is reported in the IMU sensor's own frame, and the mapping between its axes and the hand structure will be defined in a future hand model update.
sub = hand.imu().subscribe() # → Subscription[ImuData]
s = await sub.recv_async()
a, g = s.linear_acceleration, s.angular_velocity # m/s² and rad/s
print(s.header.seq, a.x, a.y, a.z, g.x, g.y, g.z)
sub.close()ImuData is a cross-device shared type. See the Wuji SDK data reference for field definitions.
Joint Traversal
JointHandle provides label / index (for indexing into 20-element return arrays such as effort_limit() / mit_params()), plus single-joint resources and actions. FingerHandle traverses joints by finger.
| Method | Return type | Description |
|---|---|---|
hand.joints() | list[JointHandle] | All 20 joints |
hand.fingers() | list[FingerHandle] | All 5 fingers |
for joint in hand.joints():
print(joint.label, joint.index) # e.g. "thumb_S1" 0
for finger in hand.fingers():
for joint in finger.joints(): # 4 joints per finger, in S1..S4 order
print(joint.label) # "thumb_S1", "thumb_S2", ...JointHandle
Joint handle: provides label / index, single-joint resources, and single-joint actions.
| Member | Type | Description |
|---|---|---|
label | str | Property. Joint label, format {finger}_S{1..4}, e.g. "thumb_S1", "pinky_S4" |
index | int | Property. Global index (0–19) |
effort_limit() | Resource (SET/GET) | Single-joint torque cap (A) |
error_code() | Resource (GET) | The joint's current error code — decode with describe_error(). Returns an integer, so format it as f"0x{code:04X}" if you print it directly |
status_word() | Resource (GET) | The joint's status word |
enable() / disable() | Action | Enable / disable this joint |
clear_fault() | Action | Clear this joint's faults |
set_origin() / clear_origin() | Action | Set / clear this joint's user origin |
FingerHandle
Finger handle: traverse the 4 joints of this finger.
| Method | Return type | Description |
|---|---|---|
joints() | list[JointHandle] | The 4 joints of this finger, in S1..S4 order |
Data Types
JointDiagnosticsFrame / JointDiagnosticsEntry
Joint diagnostics frame: frame type of the hand.joint_diagnostics().subscribe() stream. Variable-length, online joints only.
class JointDiagnosticsFrame:
header: FrameHeader # seq / timestamp_us / frame_id
num_joints: int
joints: list[JointDiagnosticsEntry]
comm: Hand2CommSummary # frame-level comm health summary, 1 Hz refresh (see Hand2CommSummary)
class JointDiagnosticsEntry:
nid: int # node ID — identifies the joint across frames
status_word: StatusWord # status word
current: float # current (A)
vbus_v_fb: float # bus voltage (V)
mcu_temp_c_fb: float # MCU temperature (°C)
error_code_current: int # current error code (integer value, shown in hex when the entry is printed) — decode with describe_error()
comm_response_rate_pct: int # bus response rate over the last second, 0–100
comm_timeout_total: int # cumulative bus timeouts for this jointMitParam
MIT parameters: element of mit_params().get().
class MitParam:
kp: float
kd: floatset accepts a single (kp, kd) (applied to all joints) or a list of 20 (per joint). get returns 20 entries — offline joints are None.
JointStateFrame
Joint state frame: return value of hand.joint_states().subscribe().recv(). Variable-length, online joints only.
| Attribute | Type | Description |
|---|---|---|
header | FrameHeader | seq / timestamp_us / frame_id |
num_joints | int | Joint count in this frame |
joints | list[JointStateEntry] | Joint state array |
JointStateEntry
Single-joint state.
| Attribute | Type | Description |
|---|---|---|
nid | int | Node ID — identifies the joint across frames |
position | float | Position (rad, joint-side) |
velocity | float | Velocity (rad/s) |
effort | float | Torque (A) |
JointCommand
Joint command: element of the joint_command().publish().send() argument.
class JointCommand:
position: float # position (rad)
velocity: float # velocity (rad/s)
effort: float # torque feedforward (A)HandCommunicationDiagnostics
Communication diagnostics data.
class HandCommunicationDiagnostics:
fingers: list[FingerCommunicationDiagnostics] # length 5
class FingerCommunicationDiagnostics:
tx_frame_total: int
rx_frame_total: int
tx_kbps: int
rx_kbps: int
error_per_sec: int
crc_error_total: int
frame_format_error_total: int
uart_hw_error_total: int
transfer_stats: list[TransferStats]
nodes: list[NodeDiagnostics] # per-node: online / ms_since_last_response / response_rate_pctFingertipSensorInfo / FingertipSensorData
Fingertip sensor metadata and data frames: return value of get_fingertip_info(finger) and frame type of the fingertip_{finger}_data subscriptions.
class FingertipSensorInfo:
header: FrameHeader # seq / timestamp_us / frame_id
digest: int # CRC32 of the info content — binds data frames to this info
model: str # sensor model string (may be empty)
device_type: int # sensor type enum
rate_hz: float # native output rate of the data stream (Hz, currently 100) — unchanged when the push rate is lowered
format: str # JSON: data-frame layout (point_count / point_stride / point_fields / aggregate_fields / encoding)
class FingertipSensorData:
header: FrameHeader
info_digest: int # digest of the info in effect — on mismatch, re-fetch the info
data: list[int] # pure value payload, interpreted per info.formatTactileStatus / TactileType / TactileState
Tactile status: return value of hand.tactile_status(finger).
class TactileStatus:
model: TactileType # Standard (34-point fingertip) or Thumb (40-point thumb)
state: TactileState # Ready or CalibratingEnum members compare equal to the firmware integer, for example TactileState.Calibrating == 1.
StatusWord
Status word: the type of the status_word field on JointDiagnosticsEntry, the public status view exposed by the SDK.
| Field | Type | Description |
|---|---|---|
ext_state | int | Extended state value (such as Init / Ready / Enabled / Stopped) |
ext_state_name | str | Semantic name of ext_state |
position_limit_active | bool | Position limit triggered |
velocity_limit_active | bool | Velocity limit triggered |
current_limit_active | bool | Current limit triggered |
Hand2CommSummary
Frame-level communication health summary: the type of the comm field on JointDiagnosticsFrame, combining the device-internal bus view with the SDK-local end-to-end (Ethernet) view. The device-internal part comes from a snapshot the SDK refreshes automatically at 1 Hz — no polling required.
| Field | Type | Description |
|---|---|---|
age_ms | int | Age of the device-internal snapshot in milliseconds. 65535 = never sampled successfully (the device-internal fields are zeros). 65534 = saturated, the snapshot is 65.5 s or older |
tactile_online_mask | int | Fingertip tactile online bitmap, bit0 = thumb … bit4 = pinky |
e2e_received | int | Frames received across all subscribed streams (cumulative) |
e2e_lost | int | Frames lost on the Ethernet segment, detected via sequence gaps (cumulative) |
e2e_reordered | int | Late/reordered frames (cumulative) |
e2e_duplicates | int | Duplicate frames (cumulative) |
e2e_window_loss_x100 | int | Loss rate over the last second, in 0.01% units |
rpc_total | int | Requests sent (cumulative). This field is 0 when RPC statistics aren't available for the current connection. In that case, rpc_retries and rpc_timeouts don't represent the actual retry or timeout counts |
rpc_retries | int | Request retransmissions (cumulative) |
rpc_timeouts | int | Requests that ultimately timed out (cumulative) |
comm_get_failures | int | Failed internal snapshot refreshes (cumulative) |
tactile_response_rate_pct | list[int] | Per-finger tactile (node 5) bus response rate, 0–100, index 0 = thumb … 4 = pinky |
tactile_timeout_total | list[int] | Per-finger tactile cumulative bus timeouts, same index order |
sdk_dropped | int | Frames dropped inside the SDK because one of its in-process hops fell behind (cumulative) — either your subscription consumer or the internal stream handler. Distinct from e2e_lost, which is network loss |
Reading the device-internal fields. age_ms, tactile_* and the per-joint
comm_* fields all come from one snapshot the SDK refreshes at 1 Hz. The refresh is
deliberately skipped while a firmware upgrade or a tactile calibration is running, so a
briefly stale snapshot is expected rather than a fault. Tell the two cases apart with
comm_get_failures: a growing age_ms while comm_get_failures stays flat means the
refresh is being skipped on purpose, while comm_get_failures increasing means the refresh
itself is failing.
Counters saturate, they do not wrap. Every counter in this table stops at its
maximum (65535 for the 16-bit ones — e2e_reordered, e2e_duplicates,
rpc_retries, rpc_timeouts, comm_get_failures — and 4294967295 for the 32-bit ones)
and stays there. On a 1 kHz stream a long-running session can reach those ceilings, so
read these as deltas between two samples rather than as absolute lifetime totals.
Joint Numbering
| Finger | S1 | S2 | S3 | S4 |
|---|---|---|---|---|
| thumb | 0 | 1 | 2 | 3 |
| index | 4 | 5 | 6 | 7 |
| middle | 8 | 9 | 10 | 11 |
| ring | 12 | 13 | 14 | 15 |
| pinky | 16 | 17 | 18 | 19 |
Label format: {finger}_S{1..4} (e.g. thumb_S1, pinky_S4).
Exception Handling
All SDK operation errors raise a unified WujiException whose message carries an error-type prefix (Disconnected, Timeout, PathNotFound, SchemaMismatch, SerializeError, …). Use try / except as needed.
from wuji_sdk import WujiException
try:
hand.joint_states().subscribe()
except WujiException as e:
print(f"SDK error: {e}")