Move a behavior to a different hand
The task may stay the same while geometry, sensing, actions and dynamics all change. Treat transfer as a sequence of testable adaptations.
Locate the gaps between a research policy and a target hand.
Define exactly what is transferred
A human pose, a grasp target, a learned visual representation and a complete policy are different transferable objects. State what stays fixed and what is retrained or recalibrated. A new wrist frame alone can break a policy even if the nominal number of joints is unchanged.
Map names and semantics before numbers
Joint order can differ between URDF parsers, optimizers, simulators and drivers. Dex Retargeting explicitly warns about this. Map by joint name, verify units and laterality, then validate output semantics. A name map fixes ordering; it does not solve different kinematics or change position targets into torque commands.
Retarget task geometry under constraints
Match fingertip positions or relative geometry in a consistent frame, while respecting the target hand’s reachable set, coupling and joint limits. Different thumbs and finger lengths may make exact pose matching impossible. Record the residual and identify which contacts matter for the task.
Treat visual adaptation as a separate experiment
DexUMI addresses both kinematic and visual embodiment gaps, including processing human-hand imagery. Its demonstrated hand setups do not establish universal hand compatibility. Visually plausible transformed data can still contain wrong contacts, occlusions or timing. Compare the raw and transformed evidence.
Test dynamics and sensing after geometry
A collision-free posture is not a stable grasp. Tendon coupling, friction, backlash, contact stiffness and controller delay affect execution. DextrAH provides a concrete simulation-to-vision distillation workflow. Its teacher’s privileged simulator state must not be silently available to a deployed student.
Write a transfer report with a narrow claim
Hold out target-hand tasks and conditions, compare a target-hand baseline, and report what target-hand data were used. Reserve “zero-shot” for the exact axis on which no adaptation was performed. Generalization to new objects is not the same claim as transfer to new hardware.
Work through the mechanics
Define a staged cross-hand adaptation with explicit evidence at each boundary.
Prerequisites: Forward kinematics, reference frames and the action-interface lesson.
A correct shape can still command the wrong finger
Consider an invented five-coordinate teaching hand, with one curl coordinate per finger. The optimizer outputs [thumb, index, middle, ring, little], while a downstream interface expects [index, thumb, middle, ring, little]. The array shape and numerical range remain valid when these two coordinates are swapped. Only a name-aware contract check exposes the error.
Dex Retargeting warns that URDF parsers may enumerate joints differently. Create an explicit source-to-target index map, reject duplicate or missing names, and verify units, control mode and laterality separately. Do not fall back to positional copying when a name is absent. This is an offline data transformation; it is not a hardware execution command.
The calculator below demonstrates a permutation for the same abstract joints. It cannot retarget a different mechanism. A target hand with coupled tendons, different axes or different fingertip geometry requires an embodiment-specific mapping and kinematic validation after the names are aligned.
def reorder(values, source_names, target_names):
if len(values) != len(source_names):
raise ValueError("value/name length mismatch")
if len(set(source_names)) != len(source_names):
raise ValueError("duplicate source joint")
if len(set(target_names)) != len(target_names):
raise ValueError("duplicate target joint")
if set(source_names) != set(target_names):
raise ValueError("different joint sets need an explicit adapter")
by_name = dict(zip(source_names, values))
return [by_name[name] for name in target_names]Find the silent finger swap
Synthetic teaching example · calculations only. No inference, simulation or hardware execution.
| Target joint | Raw copy (rad) | Name-aware (rad) |
|---|---|---|
| index | 0.20 | 0.80 · reordered |
| thumb | 0.80 | 0.20 · reordered |
| middle | 0.40 | 0.40 |
| ring | 0.30 | 0.30 |
| little | 0.10 | 0.10 |
Source order: thumb, index, middle, ring, little. Target order swaps index and thumb. One toy curl coordinate per finger; no device is connected.
Read the primary work: Dex Retargeting: joint names and optimization ↗
Transfer contact objectives, then audit the residual
For cross-hand geometry, use task-relevant quantities such as fingertip-object distances, thumb opposition or a wrist-relative grasp frame. Keep the targets and forward kinematics in the same coordinate frame. A target posture may be unreachable; report that failure rather than silently treating a clipped optimizer output as a successful match.
A useful geometric diagnostic is mean fingertip error in millimeters, alongside per-finger errors and the object scale. Normalizing by a stated object dimension helps compare some tasks, but it does not create a universal pass threshold. A 3 mm error can be harmless on a large object and decisive for a small edge pinch. Contact stability still depends on forces, friction and compliance.
DexUMI is instructive because it treats the kinematic gap and the visual gap separately. Inpainted or transformed images can reduce an appearance mismatch, yet require checks for altered object boundaries and occluded contacts. Compare geometry-only adaptation with geometry-plus-visual adaptation under matched conditions. Do not attribute both effects to a single opaque “transfer” score.
e_tip = (1 / N) Σᵢ ‖pᵢ(target hand) − pᵢ(reference)‖₂ e_relative = e_tip / declared object length A low geometric error does not establish stable contact.
Read the primary work: Dex Retargeting: joint names and optimization ↗ · DexUMI: kinematic and visual domain gaps ↗
Separate hardware transfer from task generalization
DextrAH documents privileged reinforcement-learning training followed by distillation toward camera-based observations in a specified hand-arm environment. Privileged state can help a teacher learn, but the student must be evaluated with only its allowed deployment observations. A student that still receives simulator object pose during testing has not demonstrated the intended sensing transfer.
DexGraspVLA provides a concrete hierarchy: a high-level vision-language planner and a diffusion-based low-level controller. Its released example uses a 13-coordinate arm-and-hand action/state representation, plus specified camera inputs. That combined vector must not be described as 13 independently controlled finger joints, or as a ready adapter for every commercial hand.
Build a staged evidence ladder: offline schema round trip; geometric replay in the chosen simulator; task tests with the actual observation set; then documented target-hardware trials under its supported setup. Compare to a target-hand baseline using the same data budget. Record which stage has actually run and which remains planned. Object-generalization claims from one setup do not establish new-hand transfer.
| Boundary | Useful check | Still unproven |
|---|---|---|
| Data → interface | Names, units, frames, normalization and timestamp contract | Feasible motion and contact |
| Geometry → simulated behavior | Reachability, collisions and task outcomes | Real dynamics, calibration and sensor noise |
| Teacher → student | Only allowed observations at test time | Reliability on the physical target hand |
| Source hand → target hand | Held-out target conditions and matched baseline | Other hardware or unseen tasks outside that evaluation |
Read the primary work: DextrAH: privileged training and visual distillation ↗ · DexGraspVLA: the released action and observation schema ↗ · Dex Retargeting: joint names and optimization ↗
Write an honest transfer claim
A paper succeeds on unseen objects with one hand. You reorder its action vector for a second hand and obtain small fingertip errors in simulation. Can you claim zero-shot transfer? Write the result you can support and the next two checks.
Reveal the worked answer
You can report an offline interface/geometry adaptation under the specified simulation conditions, with the actual residuals. The original unseen-object result does not prove transfer to different hardware. Next verify feasible closed-loop task behavior with the permitted sensor inputs, then evaluate the target hardware against a matched baseline and state all calibration or target-data adaptation. Do not label unrun stages as results.
Self-review checklist
- Separates unseen-object generalization from unseen-hand transfer.
- Names the level of evidence actually obtained.
- Checks sensing, dynamics and closed-loop task outcomes beyond pose error.
- States adaptation data and unrun stages explicitly.
Original Dexhands teaching examples. This is a self-study rubric, not automated grading or certification. Research sources reviewed September 25, 2026; examples do not report experiments run by Dexhands.
Try it yourself
Your practical task
- Compare the raw copy and the name-aware mapping in the workshop.
- Write a geometry test and a closed-loop test separately.
- List which evidence stages have run and which are planned.
What to produce: A staged target-hand adaptation plan with explicit evidence limits.
Check your understanding
Original sources & next steps
Dex Retargeting: joint names and optimization ↗DexUMI: kinematic and visual domain gaps ↗DextrAH: privileged training and visual distillation ↗DexGraspVLA: the released action and observation schema ↗Original Dexhands teaching material. Lesson and linked references reviewed 2026-09-25. Research links are not endorsements or evidence of hardware compatibility.