Skip to content

0.1.0

Latest

Choose a tag to compare

@AndrejOrsula AndrejOrsula released this 14 Oct 04:32
· 18 commits to main since this release
c6130b7

Summary

1. Tasks and Assets

New Tasks

Task IDs Description
spacewalk, spacewalk_eva, spacewalk_iva Humanoid (default Unitree G1) follows a waypoint path outside the ISS and Gateway (EVA) or inside the Kibo module (IVA)
moonfall Heavy Peregrine lander with onboard event camera: deployment stage, then several landing phases
terrain_landscaping, terrain_landscaping_crater, terrain_landscaping_mound RaphRover with RaphShovel on a regolith bed tasked with moving granular particles around; no trained policy is included
formation_following, formation_following_orbital, formation_following_decentralized Three Cadre rovers hold a formation around a moving center under one shared policy; orbital: CubeSats in microgravity; decentralized: per-rover policy views
beam_transport Two Spot robots carry a rigid beam to a target
traversal Articulated rover with actuated steering and suspension drives a straight line at a target velocity over challenging terrain
aerial_navigation Ingenuity flies to a moving waypoint, keeping clearance above the terrain
flight Ingenuity ascent, transfer, landing
modular_assembly Arm assembling modular elements (e.g. three LEGO blocks)
lunalab, zerog_lab Inspired by LunaLab and Zero-G Lab at the University of Luxembourg; not validated against the real facilities
excavation_demo Manual-control excavation with a Kinova Gen3 N7 arm

Example CLI commands:

srb ls env
srb ls --live env
srb agent teleop -e spacewalk_iva
srb agent zero   -e spacewalk env.mode=iva env.use_waypoint_path=false
srb agent zero --headless -e moonfall --record-events 'env.active_sensors=[proprioception,contact,transform,events]'
srb agent teleop -e terrain_landscaping_crater env.num_envs=1
srb agent rand   -e formation_following
srb agent zero   -e formation_following_orbital
srb agent zero   -e beam_transport env.beam_kinematic_mount_enabled=false
srb agent teleop -e traversal
srb agent teleop -e aerial_navigation
srb agent rand   -e flight env.landing_only=true
srb agent teleop --env excavation_demo env.scenery=lunalab env.demo_regolith_quality=visual
srb agent teleop -e modular_assembly
srb agent teleop -e lunalab
srb agent teleop -e zerog_lab

Updated Rewards

Every task has a new env.reward block with terms configurable from CLI. Existing tasks were re-tuned, so results are not directly comparable with 0.0.6.

srb agent teleop -e landing env.reward.penalty_crash.weight=-50
srb agent teleop -e landing 'env.active_sensors=[proprioception,contact,transform,events]'
srb agent zero   -e waypoint_navigation env.sim_to_real_profile=waypoint_navigation_synthetic_v1
srb agent teleop -e orbital_evasion env.num_obstacles=8 env.obstacle_scale=3.0

New Assets

Kind Assets
Ground robots cadre, husky, nifty, raph_rover
Aquatic robots heron
Orbital robots floating_platform, int_ball1, int_ball2
Multi-arm robots levion_arms, zeroglab_arms
Tools and payloads raph_shovel, husky_payload_lunalab
Objects beam, bolt_m8_tight, nut_m8_tight, projector_light, lego_block_{1x1,1x2,1x3,1x4,2x2,2x3,2x4}, random_lego_block
Sceneries iss_gateway_scenery, kibou, zeroglab, plane, terrain_generator
Removed objects bag, juggling_ball
srb ls robot
srb agent teleop -e _ground_manipulation env.robot=husky+ur10+robotiq_hand_e

2. Environment Features

  • Sensors: env.active_sensors selects categories proprioception, contact, transform, visual, tactile, terrain, event_camera (alias events), custom; default [proprioception,contact,transform]
  • Observations: --obs {state,visual} for teleop, collect, train, eval; --obs visual fails preflight in 0.1.0 because hyperparams/visual_profiles.yaml does not ship
  • Actions: named per-robot action modes, selected with env.robot.action_mode; new ones include smooth_osc, drive_filtered, and thruster control. smooth_osc and the Kortex interface share rate, critically damped, and jerk filters; drive_filtered and ros_cmd_vel share rate and critically damped filters (ros_cmd_vel only with filter_enabled, off by default). The environment's srb_spec.action describes the action layout
  • Wrappers: per-run observation stacking and action smoothing
  • Privileged observations: env.expose_privileged=true adds a state_privileged group to excavation (on by default in excavation_demo), and each RL preset decides which networks see it. onnx_export_input_whitelist() and assert_privileged_absent_from_onnx() in srb.core.env check that it stays out of exported ONNX policies; exporters do not call them
  • Randomization: events carry a category (visual, physics, task, sensor, lifecycle, or the default untagged)
  • Debugging: env.hud=true (in-simulation HUD), --snapshot (save one PNG after reset, then exit), --enable_cameras (render cameras without recording video)
  • Teleoperation: gamepad uses Isaac Lab's native driver instead of ROS; SpaceMouse discards stale motion after failed reads or 0.5 s without input; --no-invert_controls disables control inversion
srb agent rand   -e _manipulation env.robot.action_mode=joint_pos_rel
srb agent train  --algo sbx_ppo -e landing env.observation_stacking.enable=true env.observation_stacking.n=4
srb agent teleop -e landing env.action_smoothing.enable=true
srb agent teleop -e waypoint_navigation env.hud=true
srb agent zero   -e excavation_demo --headless --snapshot out.png --snapshot-warmup 32

Scene, Camera, and Physics Settings

New env.* settings, except delays:

  • Cameras: env.camera_resolution (default [64,64]), env.camera_data_types (default [rgb,depth]), env.camera_clipping_range (near and far planes), env.camera_record=true (per-episode videos to env.camera_record_dir)
  • Regolith particles (docs/src/config/particles.md): env.particles_spawner (pile shape grid, pyramid, crater_pile, or mound_pile); env.particles_density, env.particles_friction, env.particles_damping (value or [min,max] range); env.particles_freeze=true (freeze once settled); env.particles_conform_to_terrain=true (snap onto terrain); env.particles_pile_center_jitter (shift pile each reset)
  • Multi-robot layout: env.robot_spacing, env.robot_positions, env.robot_spawn_height
  • Scenery and lighting: env.sunlight_rpy (fixed sun orientation in degrees; conflicts with env.ephemeris), env.scenery_flat_area_size (flattened central area of procedural terrain)
  • Delays (existed in 0.0.6, now validated): env.action_delay_steps, env.observation_delay_steps (fixed value or [min,max] range)
srb agent teleop -e peg_in_hole 'env.active_sensors=[proprioception,contact,transform,visual]' 'env.camera_resolution=[128,128]' env.camera_record=true
srb agent zero   -e excavation --headless env.particles_spawner=pyramid env.particles_freeze=true --snapshot out.png
srb agent teleop -e _ground 'env.sunlight_rpy=[0,30,90]' env.scenery_flat_area_size=20
srb agent rand   -e waypoint_navigation 'env.action_delay_steps=[0,2]' 'env.observation_delay_steps=[0,2]'

3. Learning

Algorithms

New: tdmpc2, rsl_rl_ppo, Robomimic imitation learning.

  • RL (27): dreamer, tdmpc2, sb3_{a2c,ars,crossq,ddpg,ppo,ppo_lstm,sac,td3,tqc,trpo}, sbx_{crossq,ddpg,ppo,sac,td3,tqc}, rsl_rl_ppo, skrl_{a2c,ddpg,ppo,ppo_rnn,rpo,sac,td3,trpo}
  • Imitation learning (11): robomimic_{bc,bc_rnn,bc_transformer,bcq,cql,iql,td3_bc,gl,hbc,iris,diffusion_policy}

train and tune accept RL only, learn imitation learning only, and eval, collect, teleop both. Default hyperparameters changed (notably skrl sac, td3, ddpg, ppo), so results are not directly comparable with 0.0.6.

Dreamer

Dreamer trains in worker processes and can save its replay buffer for resumable runs.

srb agent train --headless --algo dreamer --env waypoint_navigation env.robot=leo_rover env.num_envs=1 env.domain=earth env.scenery=oberpfaffenhofen
srb agent train --headless --algo dreamer --env landing env.num_envs=1 agent.replay.save=true agent.replay.save_wait=true
srb agent train --headless --algo dreamer --env landing --continue_training

TD-MPC2

No --continue_training; --model imports weights only. Parallel environments on tasks with early termination are supported but not yet validated. Terrain-landscaping tasks are rejected.

srb agent train --headless --algo tdmpc2 --env landing env.num_envs=16
SRB_EVAL_EPISODES=100 SRB_EVAL_OUT=/tmp/eval.json srb agent eval --algo tdmpc2 --env landing env.num_envs=16 agent.mpc=false

Data Collection and Robomimic Offline Learning

srb agent collect --collector {human,policy,random,zero} records demonstrations with env.num_envs=1, defaulting to policy when --algo or --model is given, else human (teleoperation); policy collection and eval load the latest trained run of --algo for the task unless --model PATH is given. Flag --success_only discards failed episodes and needs a task that reports success (excavation, modular_assembly, beam_transport, flight, formation_following, terrain_landscaping, spacewalk); autonomous collectors give up after 3 × --num_episodes attempts. srb agent learn trains offline from --dataset without Isaac Sim.

srb agent collect --env peg_in_hole --num_episodes 50 env.num_envs=1
srb agent collect --env peg_in_hole --teleop_device ros --num_episodes 50 env.num_envs=1
srb agent collect --env peg_in_hole --algo sb3_ppo --num_episodes 200 env.num_envs=1
srb agent collect --env excavation --collector zero --num_episodes 5 --horizon 200 env.num_envs=1
srb agent learn --algo robomimic_bc_rnn --env peg_in_hole --dataset /path/to/demos.hdf5 --num_epochs 500 --normalize_actions
srb agent eval  --algo robomimic_bc_rnn --env peg_in_hole env.num_envs=1

Demonstration-Prefilled RL

skrl_sac, skrl_ddpg, skrl_td3, and dreamer can prefill the replay buffer from demonstrations.

srb agent train --algo skrl_sac --env peg_in_hole --demo-dataset /data/peg_in_hole_canonical --demo-max-transitions 1000 env.num_envs=1 agent.memory.memory_size=8192

Optuna Hyperparameter Tuning

srb agent tune subcommands: run, status, export, repair, baseline, check, validate. Search spaces: skrl ppo, ppo_rnn, sac; sb3 and sbx ppo, sac, tqc, td3, crossq; rsl_rl_ppo, dreamer, tdmpc2. Other RL algorithms fail at runtime without a search space. Overrides use --pin KEY=VALUE, and the environment count --trial-num-envs. tune export writes into hyperparams/.

srb agent tune check  --env sample_collection --algo skrl_ppo --trial-steps 10_000_000 --fps 50000
srb agent tune run    --env sample_collection --algo skrl_ppo --trial-steps 19_968_000 --n-trials 40 --n-workers 2 --gpus 0,1
srb agent tune export --study srb-skrl_ppo-state-sample_collection-v1 --to task --dry-run

Model Artifacts and Publishing

srb model push publishes artifacts to the Hugging Face Hub; eval and collect infer --algo from an artifact.

srb agent train --algo rsl_rl_ppo --env excavation --model-artifact-output /models/srb_excavation_rsl_rl_ppo --model-artifact-task excavation
srb model from-run logs/excavation/rsl_rl_ppo/RUN_DIR --dataset /data/srb_excavation_policy --output /models/m
srb model validate /models/m --strict
srb model push /models/m --dry-run
srb agent eval    --env excavation --model /models/m
srb agent collect --env excavation --model /models/m env.num_envs=1

4. Sim-to-Real

Generating Real-Robot Environments

srb real_agent gen (alias sim2real_gen) generates real-robot environments from simulated tasks; --env ALL generates all cached environments.

srb real_agent gen --env waypoint_navigation env.robot=leo_rover --hardware ros_cmd_vel ros_tf ros_mw
srb real_agent gen --env terrain_landscaping_crater --check   # print a diff against the checked-in module; fail on drift without writing

Hardware interfaces (15):

  • Generic ROS 2: ros_cmd_vel, ros_tf, ros_imu, ros_image, ros_mw
  • Perception and targets: ros_depth_heightmap, ros_optitrack, ros_target_patch, ros_scoop_target, ros_formation_following_observation
  • Manipulators: ros_kortex_cartesian, ros_kortex_joint_vel, moveit_servo, moveit_gripper
  • Testing: dummy

Deployment

Runtime zero, rand, teleop, ros, train, eval, collect operate on the shipped real-robot environments excavation, formation_following, terrain_landscaping, terrain_landscaping_crater, and waypoint_navigation. Real-robot teleoperation runs without Isaac Sim and supports only the ros (default), spacemouse (needs srb[hw], included in all), and haptic devices. Only dreamer is allowed to train on real hardware for now (experimental continual learning on the robot), as the rest is not validated.

srb real_agent teleop  --env waypoint_navigation --teleop_device spacemouse
srb real_agent eval    --env waypoint_navigation --algo dreamer   # latest trained run, or pass --model
srb real_agent collect --env waypoint_navigation --collector random --num_episodes 5
srb real_agent train   --env waypoint_navigation --algo dreamer --continue_training

Validation and Reporting

srb real_agent validate runs a profile from hyperparams/validation/ on the robot and reports pass or fail. Robot runs need --checkpoint unless the profile sets one (the shipped profiles use REQUIRED_AT_RUNTIME); --dry-run needs none, checks only the spec and session setup, and exits 4 once the spec loads.

srb real_agent validate --spec hyperparams/validation/formation_following.yaml --dry-run
srb real_agent validate --spec hyperparams/validation/formation_following.yaml --checkpoint CHECKPOINT --episodes 10
srb real_agent rebuild-summary SESSION_DIR
srb real_agent replay-twin --session SESSION_DIR --baseline BASELINE.rrd

ROS 2 Interface

  • New action/raw topic (std_msgs/Float64MultiArray) carries the flat action vector; single-environment topics drop the env0/ prefix.
  • QoS is best-effort with depth 1, commands expire after 0.5 s, and resets apply after the next step.
  • --ros_sync waits for a fresh action at every step after the first; --ros_action_timeout limits that wait in simulation and sets the startup wait and command lifetime for real_agent ros; --ros_unthrottled removes simulation pacing but not the RealEnv action rate.
  • By default, real_agent ros waits 5 s for the first message; it aborts on missing or invalid input instead of sending zero actions.
srb agent ros -e waypoint_navigation --ros_sync --ros_action_timeout 2.0         # simulation
srb real_agent ros --env waypoint_navigation --ros_sync --ros_action_timeout 2.0  # real robot

5. Terrain, Ephemeris, Sensors, and Data

Real Terrain from Orbital Data

Body Elevation Imagery
Moon LOLA DEMs (PGDA), LROC NAC DTMs, ShadowCam DTMs, shape-from-shading DEMs LROC NAC and WAC mosaics, ShadowCam mosaics
Mars MRO HiRISE DTMs, CTX DEMs HiRISE orthoimages
Earth (analog sites) Copernicus GLO-30, USGS 3DEP —

Curated sites (lunar_site_apollo17, martian_site_jezero, …): apollo15 and apollo17 (LROC NAC DTMs), jezero (CTX DEM, 20 m), jezero_v2 (HiRISE DTM and orthoimage, 1 m). Any latitude/longitude region up to 8 km covered by these sources can also be fetched, or your own DEM baked. A source layer whose estimated fetch exceeds 1 GiB needs SRB_TERRAIN_LARGE_OK=1 and catalog approval (allow_large_download: true); v1 curated sites (apollo15, apollo17, jezero) do not need the variable. Large downloads can fail while NASA's server is unstable (see Known Issues). srb asset stack previews offline which sources would be fused; --num-locations downloads several nearby regions.

# List, download, and generate a new terrain based on region latitude/longitude and size
srb asset list terrain --body moon
srb asset download terrain apollo17 --lod medium
SRB_TERRAIN_LARGE_OK=1 srb asset download terrain --body moon --lat -89.66 --lon 0.0 --size-m 2000
SRB_TERRAIN_LARGE_OK=1 srb asset download terrain --lat -89.66 --lon 0.0 --size-m 2000 --num-locations 8 --location-spread-m 3000
srb asset stack --body moon --lat -89.46 --lon 137.3 --size-m 2000 --json
srb asset bake terrain --dem ./my_dem.tif --body moon --site-name my_site --attribution NASA/LROC
srb asset doctor terrain jezero --body mars --lod low   # exits 1 and prints the fix if the site is not baked

# Use real terrain in a task
srb agent teleop -e _ground env.domain=mars env.scenery=martian_site_jezero
SRB_TERRAIN_LARGE_OK=1 srb agent teleop -e _ground env.domain=moon env.scenery.body=moon env.scenery.lat_deg=-89.66 env.scenery.lon_deg=0.0 env.scenery.size_m=2000

Ephemeris

Sun and Earth positions come from real ephemerides for a given site and epoch. Opt-in env.ephemeris options:

  • expose_observations=true adds obs["ephemeris"], but crashes in 0.1.0 (see Known Issues)
  • eclipse=true dims the Sun when a body such as Earth blocks it; drive_skydome=true (with drive_sunlight=true, surface domains only) rotates the skydome with the Sun
  • Orbital tasks can use a circular-orbit observer, configured in Python only
srb ephemeris download
srb ephemeris info --body moon --lat -89.66 --lon 0.0 --utc 2026-11-01T12:00:00 --bodies sun,earth
srb agent teleop -e _ground env.ephemeris.enabled=true env.ephemeris.epoch.utc=2026-11-01T12:00:00
srb agent teleop -e _ground env.ephemeris.enabled=true env.ephemeris.eclipse=true
srb ephemeris import --source pgda-lpsr --input lpsr.tif --out illumination.npz --crs "IAU_2015:30135" --center-xy 0.0,10309.969398639505 --size-m 4000 --gsd-m 5.0 --lat -89.66 --lon 0.0

Event Camera

Enable with events in env.active_sensors. The simulated camera emits raw events (x, y, timestamp, polarity) from simulator frames, modelling a Sony IMX636 whose presets set contrast thresholds, bandwidth, leak, shot noise, and refractory period. It is not yet calibrated against real recordings. SRB adds hot pixels for every backend except noop. Policies consume events as a raw list or as voxel_grid, histogram, time_surface, or stacked tensors.

Backend Preset use Setup
v2e (default) Full preset srb[event-v2e] and pinned source (see docs/src/sensors/event_camera.md)
v2e_streaming Requires shot noise 0 srb[event-v2e], pinned source, and SRB patch (python -m srb.core.sensor.event_camera.backends.v2e_install --apply)
metavision Thresholds and refractory period only srb[event-metavision] and provider installed manually
v2ce (learned model) Ignores the preset srb[event-v2ce] and provider installed manually; checkpoint from srb event-camera fetch-weights --mirror hf-andrejorsula-v2ce, set as v2ce_model_path
noop (for tests)

New commands include srb dataset export-zenodo-bundle, export-raw-evt3, validate-events reconstruction, and srb event-camera fetch-fixture (real reference recordings). RAW EVT3 and ROS bag exports reject record-events output.

srb event-camera backends --detailed

# Record in simulation
srb agent zero --headless --env velocity_tracking --record-events 'env.active_sensors=[proprioception,contact,transform,visual,events]'
srb agent collect --headless --env velocity_tracking --collector random --num_episodes 5 --horizon 50 env.num_envs=1 'env.active_sensors=[proprioception,contact,transform,visual,events]' --record-events

# Dataset tooling
srb dataset record-events --output events --backend noop --seed 42
srb dataset validate-events stats --synth synthetic.h5 --real recorded.h5 --output stats.json
srb dataset export-lerobot-v3 events/RUN_DIR --output ./bundles/lerobot
srb dataset import-raw-evt3 X.raw --output X.h5 --task landing
srb dataset export-rosbag X.h5 --output ./bag --msg-type event_camera_msgs   # needs srb[event-ros]; fails on ROS 2 Jazzy (see Known Issues)

Standard Dataset Format

srb dataset export formats: srb_jsonl, robomimic_hdf5, lerobot. push and pull use the Hugging Face Hub.

srb dataset validate /data/srb_excavation --deep --strict
srb dataset compatibility /data/srb_excavation --algo robomimic_bc --purpose train
srb dataset export /logs/excavation/collect/datasets/demos.hdf5 --output /data/srb_excavation_policy --format srb_jsonl
srb dataset replay DATASET --backend rerun --save out.rrd
srb dataset pull --repo-id org/name --dry-run

6. CLI, GUI, and Infrastructure

Commands

0.0.6 0.1.0
agent, real_agent, list/ls, gui, repl/python, docs, test agent, real_agent, list/ls, dataset, model, asset/assets, event-camera, ephemeris, cache, gui, isaac/isaacsim, doc/docs, test, clean, completions

Command modules load lazily, so srb --help starts quickly. Only agent, real_agent, and the forwarding commands (gui, isaac, test, and doc with no action, serve, or build) accept arguments after --; agent learn, agent tune, doc docgen, and doc thumbnail reject them. A missing optional feature prints the exact pip install 'srb[<extra>]' fix.

srb cache status --json --strict          # also: update, doctor, claim-clean-root
srb clean srb -y                          # targets: all | isaacsim | simforge | srb
srb isaac python -- -m pip list
srb isaac sim -- --no-window
srb doc docgen --static-only              # also: serve, build, thumbnail
srb test --tier unit                      # tiers: all | unit | workflows (workflows need Isaac Sim)
srb agent zero -e _manipulation --repl    # attach a REPL; replaces `srb repl`

Tab Completion

Tab completion is now everywhere! It completes commands, flags, task IDs, algorithms, and Hydra overrides (env.* always, agent.* once --algo is set) in Fish, Bash, and Zsh, including inside the Docker image. Override completion needs jq on the host.

srb completions generate   # build the completion cache (launches Isaac Sim)
srb completions install    # install the Fish/Bash/Zsh completion scripts

# Type a partial override, then press Tab twice:
srb agent zero -e peg_in_hole env.<Tab><Tab>
srb agent train --algo dreamer -e landing agent.replay.<Tab><Tab>

Configuration and Rendering

Hydra overrides can also come from a YAML file via --cfg. Rendering modes are now RTX presets applied after Kit starts; the old apps/rendering_modes/*.kit files are gone.

srb agent zero -e peg_in_hole --cfg ./my_overrides.yaml
srb agent zero -e peg_in_hole env.seed=1 env.seed=2 env.num_envs=4        # env.seed=2 is applied
srb agent zero -e excavation_demo --headless --rendering_mode performance

GUI

New Tasks, Session, and Metrics pages; a Configure window that previews and copies the equivalent CLI command; a reward plot from /gui/reward; a footer theme toggle and episode recording; a recovery page when the catalog cache is invalid (srb cache update refreshes it). srb gui uses a prebuilt binary from SRB_GUI_BINARY if set, or when Cargo is unavailable; otherwise it builds with cargo run. Only the rand, teleop, and zero workflows are supported.

srb gui
SRB_GUI_BINARY=/path/to/gui srb gui

Environment Variables

Variable Purpose
SRB_EVAL_EPISODES, SRB_EVAL_OUT Episode count and JSON output path for bounded eval; stochastic policies are rejected (Dreamer, TD-MPC2 unless agent.mpc=false, stochastic Robomimic checkpoints or --model artifacts); SKRL only caps its budget
SRB_CACHE_ROOT Cache location (default .cache/ in a source checkout, else $XDG_CACHE_HOME/srb or ~/.cache/srb)
SRB_TERRAIN_{AUTO_FETCH,CACHE_ROOT,CATALOG_DIR,LARGE_OK} Terrain download and cache behavior
SRB_EPHEMERIS_{AUTO_FETCH,CACHE_ROOT,HF_REPO_ID,TABLE_PRODUCT} Ephemeris download and cache behavior
SRB_REPL, SRB_GUI_BINARY (or SRB_GUI_BIN, SRB_GUI_PATH), SRB_DISABLE_ISAAC_LOG_FILTERS REPL attach, GUI binary, and unfiltered Isaac Sim logs
SRB_DISABLE_OMNI_TELEMETRY, SRB_ENABLE_OMNI_TELEMETRY Exclude Kit telemetry extensions; the second overrides the first
SRB_KORTEX_PASSWORD Kortex arm password for sim-to-real; must not be the factory default, and a password in the config is rejected
SRB_TARGET_PATCH_MAP_PATH (or SRB_TARGET_MAP_PATH) Target map for the ros_target_patch interface
SRB_ALLOW_UNVERIFIED_LANDSCAPING_CHECKPOINT Accept Dreamer landscaping checkpoints with missing metadata (not mismatched metadata)

Docker

# Run the published 0.1.0 image (~27 GB download)
SRB_DOCKER_PROFILE=simulator SRB_DOCKER_IMAGE='ghcr.io/andrejorsula/space_robotics_bench@sha256:631b2cd0c1f56ee63253c7da698db110fbf67b77294753a95cf18eb434a2cabf' ./.docker/run.bash

# Or build the image
SRB_IMAGE_NAME=space_robotics_bench ./.docker/build.bash 6.1.0

# Run a container with GPU and GUI access
SRB_DOCKER_PROFILE=simulator SRB_DOCKER_IMAGE='space_robotics_bench:6.1.0' ./.docker/run.bash

# Run one command non-interactively (CI and scripts)
DOCKER_TTY=false ./.docker/run.bash srb --help

In the simulator profile, run srb cache update once at the container prompt; the cache persists in the repository's .cache/ unless WITH_DEV_VOLUME=false. SRB_DOCKER_IMAGE must name a local image or one pinned as NAME@sha256:DIGEST, which Docker pulls on first use. SRB_DOCKER_PROFILE sets the container's host access; more than the default safe is opt-in:

Profile Grants Use for
safe (default) No network, private IPC, all Linux capabilities dropped; by default no GPU, display, or host mounts Commands without Isaac Sim, e.g. srb --help; host files only via -v
simulator GPU, X11 display, host network and IPC, Isaac Sim and SimForge cache mounts, shell history, repository read-write at /root/ws; unprivileged, no-new-privileges Isaac Sim and everyday work; edits and offline task catalog persist on the host
trusted-simulator simulator plus --privileged and the host's entire /dev Hardware passthrough only, e.g. USB devices, SpaceMouse and haptic teleoperation
ngc NVIDIA's official Isaac Sim 6.1 image (nvcr.io/nvidia/isaac-sim, pinned by digest) as UID 1234:1234, repository at /workspace; needs an RTX-class GPU Checking SRB against the upstream image; SRB tasks, ROS 2, and Rust untested there

Warning: trusted-simulator gives the container effectively full control of the host. Use it only with an image you built or one pinned by digest.

Per-profile overrides: WITH_GPU, WITH_GUI, WITH_HISTORY, WITH_DEV_VOLUME (true or false); WITH_HOST_NETWORK=false (bridged network); WITH_HOST_IPC=false (private IPC with SRB_DOCKER_SHM_SIZE of shared memory, default 8g). ngc modes: SRB_NGC_MODE=headless (default), webrtc, or gui, or the shorthand profiles ngc-headless, ngc-webrtc, ngc-gui.

SRB_DOCKER_PROFILE=simulator WITH_GUI=false ./.docker/run.bash                        # headless
SRB_DOCKER_PROFILE=simulator WITH_HOST_NETWORK=false ./.docker/run.bash               # Docker bridge network instead of host
SRB_DOCKER_PROFILE=simulator WITH_HOST_IPC=false SRB_DOCKER_SHM_SIZE=16g ./.docker/run.bash
SRB_DOCKER_PROFILE=ngc-webrtc ./.docker/run.bash                                      # NVIDIA image, WebRTC streaming

Note: The default SRB_DOCKER_PROFILE might be changed to "simulator" in the future.


Breaking Changes

CLI

  1. srb repl and srb python are removed; use --repl [ptpython|bpython] on agent commands, or srb isaac python
  2. srb ls drops -a/--show_hidden and gains --live
  3. srb agent learn no longer takes simulator flags (--headless, --video, and others) and accepts only agent.* overrides
  4. real_agent teleop no longer accepts --algo or --model. --algo values are now validated strictly
  5. srb gui -- <args> forwards the arguments to the GUI binary instead of Cargo, so srb gui -- --release no longer selects a release build

Tasks and configuration

  1. peg_in_hole_assembly and multi_peg_in_hole_assembly are renamed to peg_in_hole and multi_peg_in_hole
  2. The 26 *_visual task variants are removed; add visual to env.active_sensors instead (--obs visual needs a visual profile catalog that does not ship)
  3. env.robot.actions is replaced by env.robot.action_mode, with shorter mode names (e.g. joint_position_relative becomes joint_pos_rel)
  4. Quaternion settings such as env.tf_quat_target use xyzw order instead of wxyz
  5. With env.num_envs=1, ROS topics drop the env0/ prefix (e.g. /srb/reward)

Algorithms and platform

  1. skrl_amp and skrl_cem are removed
  2. Python 3.12, Isaac Sim 6.1.0, Isaac Lab 3.0.0, and Gymnasium 1.x are required; skrl algorithms need skrl 2.x
  3. run.bash no longer falls back to the Docker Hub image andrejorsula/space_robotics_bench; use a local image or a digest-pinned SRB_DOCKER_IMAGE

Python API

  1. Importing srb.tasks no longer registers tasks; they register once Isaac Sim is running, via register_all() or ensure_registered(). Generated sim-to-real environments register through srb.interfaces.sim_to_real.env.register_all()
  2. srb.utils.hydra moved to srb.hydra
  3. A task's extract_step_return may return an optional fifth info dict for episode diagnostics
  4. srb._typing.enum is renamed to srb._typing.enum_utils
  5. Asset modules moved into family subpackages (e.g. srb.assets.robot.mobile.ground); package-level names such as assets.LeoRover still work
  6. The Rust library and Python extension crates are removed

Commits

Added

Changed

Fixed

Removed