Add a New Inter-Chunk Fuser

Open-RAIL Add New inter_chunk_fuser Policy Development Document

inter_chunk_fuser Overview

Let the observation at time $t_1$ be $O_{t_1}$ . The VLA model predicts an $n$ -step action chunk $[a_0, a_1, \dots, a_n]_{t_1}$ conditioned on language instruction and observation.

Open-RAIL adopts an asynchronous architecture: once $\text{chunk}_{t_1}: [a_0, a_1, \dots, a_n]_{t_1}$ is available, the robot starts executing this chunk immediately. Meanwhile, the VLA model predicts the next action chunk $[a_0, a_1, \dots, a_n]_{t_2}$ based on the observation $O_{t_2}$ at time $t_2$ .

Define the VLA model inference latency as $\Delta_t$ . In general, $t_2 \ge t_1 + \Delta_t$ .

In practice, the new action chunk arrives before the old chunk finishes execution. Suppose the robot is executing the $i$ -th action of $\text{chunk}_{t_1}$ when $\text{chunk}_{t_2}$ arrives. The inter_chunk_fuser is designed to smoothly transition from the $i$ -th action of $\text{chunk}_{t_1}$ to a designated action in $\text{chunk}_{t_2}$ , mitigating abrupt jumps in action magnitude.

Input Arguments for Fuser Policy

Parameter	Description
next_action_chunk	New action sequence with dimension $d \times n$ , where $d$ = number of robot joints, $n$ = number of action steps
next_vel_chunk	Velocity sequence corresponding to next_action_chunk
next_acc_chunk	Acceleration sequence corresponding to next_action_chunk
currt_action	Current action being executed from the old action chunk
currt_vel	Velocity while executing currt_action
currt_acc	Acceleration while executing currt_action
target_chunk_index	Index of the target action to connect within next_action_chunk

Open-RAIL has implemented three built-in inter-chunk-fuser policies: search_action, min_jerk, smooth_velocity

Development Steps for Custom inter_chunk_fuser Policy (`your_fuser_method`)

Policy shorthand name: your_fuser

1. Add policy entry in inter_chunk_fuser.py

File path: ./client/core/inter_chunk_fuser.py Add a conditional branch for your_fuser inside the process function of class InterChunkFuser.

2. Implement policy logic in inter_chunk_fuser.py

File path: ./client/core/inter_chunk_fuser.py Add a member function your_fuser_method inside class InterChunkFuser to implement your smooth transition algorithm.

3. Add policy configuration in client_conf.py

File path: ./conf/client_conf.py

Create a new function your_fuser_method_config to define hyperparameters for this policy;
Instantiate your_fuser_method_config inside the get_inter_chunk_config function.

4. Add frontend configuration option

File path: ./web_client/static/modules/config.js Add the option "your_fuser" to the inter_chunk_mode list under CONFIG_SELECT_OPTIONS.

5. Verification

Start server and web_client, then visit localhost:9000. You can select the your_fuser policy from the dropdown menu at Main Parameters → Inter-Chunk Mode on the web page.