FADA | Ishayu S. Shikhare

Overview

At the LeCAR Lab, we asked a simple question about humanoid control: if a robot already knows what motion to produce, can a few minutes of its own deployment rollouts teach it how to produce that motion under new physics?

FADA (Few-Shot Domain Adaptation via Dynamics Alignment) is our answer. It is a Planner–Inverse Dynamics Model controller that specializes to a new domain from roughly two minutes of onboard rollouts — no rewards, no mocap, no data labeling, and no simulator refitting. After that update, a Unitree G1 can track a line up a narrow slope, dance with a 3.2 kg front load, and run Kung Fu on soft mats; a Booster T1 can pull a 6 kg laundry basket and circle-walk with an asymmetric arm payload. Zero-shot transfer fails at these tasks.

This is joint work with Angchen Xie, Nikhil Sobanbabu, Alan Wang, Max Simchowitz, and Guanya Shi. The paper is in submission to CoRL 2026 and was accepted at the RSS 2026 Sim2Real Workshop.

The Problem

Humanoid policies are trained in simulation, where you can fall, collide, and randomize physics at scale. On hardware, the same command no longer produces the same body motion. Terrain, payload, actuator response, and contact all shift the dynamics, and on a tightly coupled whole-body system a few centimeters of foot-placement error is enough to walk off a ramp or stall a pull.

The usual responses sit at two extremes:

Zero-shot robustness (domain randomization, history-conditioned in-context adapters) never updates weights on the target robot, so the policy stays conservative under the actual deployment condition.
Heavy target-domain learning (system ID, residual dynamics, real-world RL, full-policy finetuning) can specialize, but it wants a model to fit, a reward to optimize, or expert demonstrations to imitate.

We wanted the middle: use the robot’s own imperfect rollouts, and update only the part of the policy that actually depends on the new physics.

The Idea

Under a dynamics shift, the task intent is often still right. Walking up the ramp still means placing feet on the plank; pulling the basket still means leaning back and driving contact. What changes is the action required to realize that intent.

FADA makes that split explicit. A planner maps the command and recent observations to a short-horizon proprioceptive future — the motion the robot should produce. An inverse dynamics model (IDM) maps that future, plus recent execution history, to an action chunk. At deployment we freeze the planner and finetune only the IDM, so the robot keeps the same command-to-intent interface and learns a new plan-to-action map.

Planner–IDM architecture. The planner predicts a short-horizon proprioceptive future; the IDM maps that future and recent history to an action chunk. — The planner predicts a K-step proprioceptive future. The IDM attends to that future and to recent observation–action history, then emits an action chunk. Only the first action is executed in a receding-horizon loop.

A fixed-base arm experiment makes the split concrete. The arm tracks the same Cartesian targets while the wrist payload jumps from 0 kg to 5 kg. The planner’s predicted joint configuration barely moves (~7%), because the kinematics did not change. Finetuning only the IDM cuts tracking error by ~24%: the required correction is a structured, configuration-dependent inverse-dynamics update, not a constant action offset.

How It Works

FADA three-stage pipeline from oracle training through few-shot IDM adaptation to hardware deployment. — Source training happens in IsaacSim. Target adaptation uses a few ordinary rollouts on hardware or in a held-out simulator, with LoRA adapters on the IDM.

The pipeline has three stages.

1. Privileged oracle. We train a teacher in IsaacSim with task rewards and privileged state (contacts, terrain, actuator parameters, sampled randomization). This policy is not deployable; it exists to label good behavior.

2. Planner–IDM distillation. A student that sees only proprioception is trained with DAgger. Two losses matter. The IDM is supervised on the executed first action associated with a realized future — including suboptimal student rollouts — so the same objective later applies to imperfect target data. The planner is trained through a stop-gradient IDM, so it has to emit futures that actually produce oracle-consistent actions, not futures that merely look like oracle observations.

3. Few-shot IDM adaptation. We roll out the source student in the target domain, freeze the planner and the pretrained IDM, and train LoRA adapters (rank 8) with the same first-action inverse-dynamics loss. Locomotion uses ~2 minutes at 50 Hz (~6000 steps); whole-body tracking uses six repetitions of a ~20 s motion. Full IDM finetuning overfits this budget; LoRA does not.

Both modules are small transformers (history length 30, prediction horizon 6). The IDM decoder uses full attention over the predicted future, so later tokens can shape the deployed first action even though only that first action is supervised. A 1-step horizon is myopic; 6 steps is enough; longer horizons do not help.

Results

We evaluated on Unitree G1 (29 DoF) and Booster T1 (23 DoF), in IsaacSim-to-MuJoCo transfer and on hardware. Baselines share the same privileged oracle: a transformer DAgger student with no target update, and a co-prediction world-model-style student that we also finetune on target rollouts.

On hardware, few-shot IDM adaptation improves all five quantitative tasks:

Completion. Average success on slope traversal and basket pulling goes from 20% zero-shot to 90%. The transformer DAgger baseline finishes neither.
Tracking. Normalized velocity / MPJPE error drops 27% versus our own zero-shot student and 16% versus transformer DAgger, across grocery carrying, Kung Fu on soft mats, and T1 payload circling.
The IDM’s own action-prediction loss falls on every task, which is the mechanism: the plan-to-action map is closer to the target physics.

The same pattern holds in sim-to-sim. Averaged across five G1/T1 tasks, FADA reduces normalized error 25% versus zero-shot and 27% versus transformer DAgger. The largest gain is on T1 Falcon, a force-adaptive locomotion task where a persistent 30 N pull has to be compensated — exactly the setting where a mismatched action map hurts. Finetuning the co-prediction baseline on a future-observation loss worsens transfer: better next-state prediction is not the same as a better action generator.

Two ablations we cared about:

Removing first-action-only IDM supervision (instead supervising the full action chunk) drops zero-shot MuJoCo tracking success from 10/10 to 4/10. The training objective has to match receding-horizon execution.
The 6000-step budget is on the plateau. More target data does not keep helping; a few hundred steps is already not enough.

What this suggests

A lot of humanoid sim-to-real failure is not a failure of task intent. The robot still tries to walk the ramp, pull the basket, or hit the Kung Fu keyframes; the actions no longer induce the intended motion. If you freeze intent and realign execution from the robot’s own paired observations and actions, a few minutes of data is enough to recover high-precision whole-body skills.

That recipe is not free. The zero-shot student has to stay upright long enough to collect useful rollouts, the IDM is still trained inside a task distribution, and adaptation is proprioception-only. Those are the limitations we are working on next.

Hardware videos, charts, and an interactive arm demo are on the project website.

High-precision humanoid control is limited by target-domain dynamics mismatch, where the same control objective can induce different realized motions under changes in terrain, payload, or actuator response. Existing methods either pursue zero-shot transfer through domain randomization or in-context adaptation without target-domain specialization, or require heavy adaptation pipelines that leverage target-domain data, such as model calibration, residual learning, or policy retraining. We present FADA (Few-Shot Domain Adaptation via Dynamics Alignment), a three-stage Planner-Inverse Dynamics Model (Planner-IDM) framework for few-shot adaptation in humanoid control. FADA first trains an oracle policy with privileged information and then distills the oracle behavior into a deployable Planner-IDM student through DAgger. At deployment, FADA freezes the planner and finetunes only the IDM using approximately 2 minutes of target-domain rollouts with standard supervised learning. Rather than requiring optimal demonstrations or rewards, FADA uses the paired actions and observations observed during these rollouts as supervision, aligning the IDM’s action generation with target-domain dynamics. Experiments show that FADA outperforms both in-context and end-to-end adaptation baselines, improving task performance under dynamics shifts and enabling real humanoid robots to execute diverse high-precision whole-body tasks.